Highlights: Testing risk at the point of authority

Executive Summary

The central theme across this week's developments is that digital risk programs are increasingly being tested at the point where authority is exercised rather than where compliance is documented. Controls, policies, audits, insurance policies, and identity systems remain important. Still, the evidence suggests organizations create more value when they focus on activities that directly influence outcomes rather than activities that demonstrate oversight.

In governance and compliance, the strongest controls are not always the most widely adopted. The Association of Certified Fraud Examiners found that proactive monitoring and management review produced some of the largest reductions in fraud losses and detection times, even though many organizations continue to invest heavily in more common controls. At the same time, the Secure Controls Framework survey showed organizations struggling with growing framework requirements, audit burdens, staffing gaps, and immature AI governance. Federal procurement reform reflects a similar principle by aligning oversight requirements more closely with the actual risks and relationships being governed.

In cybersecurity, trusted infrastructure emerged as a recurring attack surface. The NVIDIA and CrowdStrike research demonstrates how AI can accelerate defensive testing and validation cycles, while the Kaspersky and BigDiskBuster examples show how administrative privileges, Group Policy, and antivirus update mechanisms can be exploited without relying on traditional malware techniques. The implication is that organizations must increasingly monitor legitimate administrative pathways as closely as they monitor malicious code. Practical steps include regularly reviewing privileged access and administrative logs, implementing real-time alerts for creating or modifying Group Policy Objects, monitoring the issuance and use of elevated credentials, and setting up automated notifications for unusual administrative activity. Additionally, deploying tools that track authorized management console use and audit critical configuration changes can provide early warning of suspicious behavior.

Audit, identity, insurance, and privacy developments point toward the same operating reality. Audit evidence is becoming continuous. AI agents are becoming identities that require governance. Insurance carriers are clarifying how autonomous-system losses fit within traditional coverage models. Privacy regulators are focusing on accountability for automated decisions, profiling, and historical data retention. In each case, organizations need clearer ownership, stronger evidence, and more direct connections between actions and accountability.

Enterprise digital-risk products reflect the same shift. Products are increasingly valued not for reporting problems after the fact, but for constraining actions in real time, governing permissions before they are exercised, and preserving evidence after decisions are made. Authorization design is becoming a control discipline rather than simply an access-management function.

The broader pattern is that digital risk is moving from governance-by-documentation to governance-by-execution. The organizations that adapt most effectively will be those that understand which actors can act, what authority they possess, what evidence is preserved, and how quickly actions can be constrained when conditions change.

Deep Research

1. GRC (Governance, Risk & Compliance)

This week's GRC sources examine whether control design reflects demonstrated risk reduction, program capacity, and the actual relationship being governed. The Association of Certified Fraud Examiners measures the controls associated with lower loss and faster detection. The Secure Controls Framework Council measures framework load, staffing, tooling, and AI governance. The Federal Acquisition Regulatory Council proposes procurement requirements that are more closely tailored to contractor relationships. Across the three sources, documentation is most useful when it directs resources toward the controls and relationships that carry the material risk.

ACFE’s Occupational Fraud 2026 study measured eighteen anti-fraud controls at 2,402 victim organizations across 143 countries and territories, with total losses of more than $3.4 billion. Among those victims, 88% had a code of conduct, 83% had an external financial-statement audit, 80% had an internal audit department, and 73% had a hotline. All eighteen controls were associated with lower losses and faster detection. The four tied to the largest reductions in median loss and duration were management review, proactive data monitoring and analysis, a code of conduct, and surprise audits. Management review, in place at 71% of victims, cut median loss from $186,000 to $84,000. Proactive data monitoring, in place at only 49%, cut median loss from $150,000 to $70,000 and cut median duration from sixteen months to nine. Surprise audits were adopted in 44% of organizations and also showed a significantly lower median loss. The almost universal controls didn't impact the losses much.

The ACFE results support a practical review of the control portfolio. Leaders can compare the controls with the strongest associations to lower loss and faster detection against current adoption, ownership, and monitoring, then use those gaps to set the next investment priority. Typically, controls that yield the highest return on investment and risk reduction include management review, proactive data monitoring and analysis, a well-enforced code of conduct, and surprise audits, as these have shown the greatest impact on reducing both loss amounts and detection times. Prioritizing these domains can help organizations focus limited resources on the controls that deliver the strongest outcomes.

The Secure Controls Framework 2026 State of the Industry Survey makes the same point at program scale. Sixty-two percent of organizations track four or more distinct laws, regulations, or frameworks at once. Nearly a third face more than ten external audits a year. Only 58% of practitioners feel confident they could withstand an unannounced audit. GRC-platform satisfaction was the lowest-rated metric in the survey, and close to half of respondents still run governance, risk, and compliance work on spreadsheets or with no dedicated tool. Mid-market firms (250 to 4,99 employees) reported the weakest executive engagement, staffing, maturity, GRC-platform satisfaction, third-party risk pressure, and Zero Trust progress among the three size bands. Fifty-six percent already use AI in compliance, GRC, or security workflows. Only 6% have a formal, externally validated AI governance program. “AI-powered” product labeling was the most distrusted marketing claim, cited by 67%.

The FAR Council’s proposed Revolutionary FAR Overhaul is the public-sector version of the same design test. It implements Executive Order 14275 and OMB Memorandum M-25-26. Written comments closed July 23, 2026. Part 4 is the section expected to significantly impact contractors: fewer identification instruments, entity-level SAM, and commercial-contract clause cleanup. Parts 1, 2, 33, 39, and 40 focus on readability, protest-process clarity, and internal government procedure. The rule is not a thicker binder. It claims the clause should match the relationship.

Sources

SCF State of the Industry Survey 2026

2. Cyber Security

This week's cyber sources examine attacks and defenses that operate through trusted infrastructure. NVIDIA and CrowdStrike test a continuous attack-and-defense workflow using specialized AI agents. Kaspersky describes extortion delivered through Active Directory Group Policy. A public proof of concept targets the process Microsoft Defender uses to install security updates. The sources place greater importance on administrative behavior, update freshness, and validated detection than on the presence of familiar security products alone.

NVIDIA and CrowdStrike evaluated an agentic offense-defense system in an isolated environment modeled on NVIDIA accelerated computing infrastructure. Nemotron models customized for cybersecurity operate inside CrowdStrike SafeMind. CrowdStrike reports that its Blue Solano defensive model was more accurate than the leading proprietary frontier model tested, at 99 percent lower cost, in CrowdStrike internal evaluations. The operating idea is a continuous loop: controlled attacks produce telemetry and ground truth, defensive agents expose coverage gaps, and the detections are retested. Traditional red-and-blue work still depends on manual handoffs. The claim is that open models plus specialized harnesses can run that loop at machine speed. Treat the published accuracy figure as a vendor evaluation until an independent test repeats it. The design question still stands: can the detection be generated, validated, and redeployed before the next seeded attack run finishes?

Kaspersky’s Global Emergency Response Team described a Payload ransomware incident at a manufacturing company in the Middle East that did not use a conventional encryptor. Attackers obtained privileged credentials, likely through phishing, logged in through standard remote-access and VPN paths, and appeared to be IT staff. They created a malicious Group Policy Object in Active Directory, locked screens, displayed ransom notes, and changed wallpapers. Kaspersky frames this as encryptionless extortion: operational disruption plus data leak leverage, without cryptographic keys. Recommended controls sit at the administrative layer: monitor GPO creation and modification, alert when a new rule is linked at the domain root, require phishing-resistant multi-factor authentication on administrative and VPN paths, and keep Domain Admin accounts off general workstations.

On September 22, The Cyber Security Hub reported that researcher Abdelhamid Naceri, also known as Nightmare Eclipse, released proof-of-concept code named BigDiskBuster. The technique targets how Microsoft Defender stages and installs security-intelligence and platform updates. If it works as described, Defender keeps running while signatures, detection logic, and engine components stop refreshing. Naceri says the issue affects currently supported Windows versions. Microsoft had not confirmed a CVE, issued a specific fix, or documented in-the-wild use at the time of the report. The honest description is a public proof of concept for update denial, not a confirmed campaign. Watch update freshness. Treat unexplained Defender update failures as possible interference, not only as a maintenance miss.

Security teams should monitor Domain Admin use, Group Policy changes, remote-access activity, and Defender update freshness as separate control signals. The incidents show that compromised credentials and administrative mechanisms can produce material disruption without introducing a conventional encryptor or visibly disabling the endpoint product. Leaders should require regular reporting on these administrative controls and notable incident trends to ensure oversight stays aligned with emerging operational risks.

Sources

3. Internal Audit, IT Audit & Continuous Control Monitoring

This week's audit sources examine how evidence, automation, and accountability change when testing becomes continuous. EY describes agents operating inside its global audit platform, with each auditor retaining responsibility for review. Research on accelerated filers describes full-population scoring that directs human attention to exceptions. Federal banking and credit-union regulators propose third-party oversight proportionate to the relationship. The consistent requirement is current evidence that a named reviewer can understand, challenge, and support.

In April 2026, EY embedded a multi-agent framework in EY Canvas and rolled it out to 130,000 assurance professionals after pilots. The platform supports more than 160,000 engagements and, EY says, processes up to 1.4 trillion journal-entry lines a year. Research agents came first in 2025. Project-management agents followed in spring 2026. Drafting and review agents arrived through the summer. EY says agentic support is expected across end-to-end audit activities by 2028. The platform includes architectural transparency and immutable audit logging. Each auditor remains the mandatory human in the loop, with professional skepticism and legal responsibility for AI-generated workpapers. The April launch sits inside a multi-year plan, announced four years earlier, to invest more than $1 billion and standardize assurance systems on EY’s private cloud. Richard Harrison of EY UK made it clear that auditors still review the work.

Hunter Gibson’s September 20 account of 42 accelerated filers describes the testing counterpart. Sixty-three percent of companies now use technology for SOX compliance, per the Protiviti Sarbanes-Oxley Compliance Survey cited in the piece. Nearly 40% fail at least one control annually. Teams using validated full-population scoring, including Isolation Forest methods, investigate flagged exceptions instead of proving clean transactions are clean. Continuous monitoring reduces routine follow-up to about 30% of prior effort and saves about 7 hours per control cycle. One illustration moves a cycle from 214 hours to 132. The limit is also called model miss, and hours drop only when review targets the exception cluster rather than the whole population.

The NCUA, Federal Reserve, FDIC, and OCC proposed 2026 third-party risk management guidance that restates a principles-based, tailored approach. Third-party relationships differ in form and in the extent of the operations they handle. The 2023 interagency guidance remains the starting point. Supervisors want practices sized to the relationship, documented, and independently reviewed. There is no single required structure. Size, complexity, risk profile, and the nature of the third party drive the design.

Audit plans can be evaluated against four elements reflected in the sources: current exceptions, supporting evidence, a named reviewer, and third-party oversight proportionate to the dependency. Without those elements, a clean historical sample provides a limited view of current operating effectiveness.

Sources

4. Identity Management / IAM

This week's identity sources extend governance beyond employees to agents, service accounts, and complex application estates. Okta and the Blueprint Alliance propose a multi-vendor architecture for agent identity and runtime control. Turnkey describes the discovery and role design needed to govern SAP within enterprise identity governance. Xiting explains the authentication layer used across SAP cloud services. The sources support a common operating model: each actor needs a distinct identity, bounded authority, traceable delegation, and an accountable owner.

On September 22, Okta announced the Blueprint Alliance with AWS, CrowdStrike, Databricks, Docker, Google Cloud, Lovable, Proofpoint, Salesforce, ServiceNow, Wiz, and Zscaler. GE Appliances and World Central Kitchen sit as strategic advisors. The Alliance expands a March 2026 blueprint into an open, multi-vendor reference architecture. Stated principles: treat every agent as a first-class identity, scope access to the task rather than standing access, keep delegation traceable, monitor runtime behavior continuously, make containment instant and reversible, and keep governance moving at AI speed. The four operating questions are: Where are my agents? What can they do? What are they doing? How do I respond? Gartner predicts that by 2028 an average global Fortune 500 enterprise will have more than 150,000 agents in use, while only 13% of organizations think they have the right AI-agent governance in place.

Turnkey’s enterprise IGA note is the SAP-specific version of that inventory problem. Platform selection should follow discovery of identity sources, applications, provisioning behavior, and the controls the organization actually needs. A connector on a vendor’s supported list does not guarantee every provisioning action, attribute, or business process. In a customized SAP landscape, an out-of-the-box integration may still need design work. Implementation can start read-only, then enable provisioning after the model is checked. The aim is an enterprise model that governs SAP properly, not a broad platform with thin SAP coverage.

Xiting’s SAP Identity Authentication Service (IAS) primer is the authentication layer under that model. SAP Cloud Identity Services, including Identity Authentication, sit on SAP Business Technology Platform. Cloud products such as S/4HANA Public Cloud, SuccessFactors, and Integrated Business Planning ship pre-configured to use IAS. You can connect other SAML 2.0 applications. IAS is commercially included with many SAP cloud and BTP scenarios. Third-party use may carry extra charges. That is useful only if the IGA layer above it can still answer who the agent is and what it may do.

Organizations can apply the Blueprint Alliance model by recording each agent or non-human identity, its accountable owner, authorized task, permitted resources, delegated authority, and containment method. That register gives identity teams a basis for access review, runtime monitoring, and incident response across human and machine actors.

For example, a global financial institution might create a centralized register listing all AI-powered trading bots, automation tools, and service accounts. Each entry would document the responsible team or manager, the specific transactions or data sets accessible to each bot, the scope of any delegated privileges, and steps to rapidly suspend or contain the agent's activity if suspicious behavior is detected. This implementation ensures every digital actor is visible, governed, and isolatable if required.

Sources

5. Cyber Insurance

This week's insurance sources examine how policy language responds when AI contributes to a cyber loss or causes harm through an authorized workflow. Beazley and CFC provide affirmative AI wording. RAND and market reporting describe insurers taking different approaches to existing and emerging AI exposure. The Gemini and Claude evaluation incidents illustrate why loss attribution can cross cyber, technology errors and omissions, directors and officers, and general liability policies. Coverage clarity increasingly depends on the event definition and the policy language that responds to it.

Beazley confirmed on September 17 that it offers affirmative AI coverage in its cyber and technology errors-and-omissions policies. Alessandro Lezzi, Group Head of Cyber Risks, said the endorsement clarifies that a cyber attack, AI-driven or otherwise, is covered. CFC upgraded financial-institution products to its full Cyber Proactive Response policy, including unlimited reinstatements; the policyholder does not have to pay any deductible before receiving a payout from the insurer; and explicitly states coverage for AI-related cyber exposures. QBE’s Serene Davis told Cybersecurity Dive that QBE treats AI as a risk amplifier, not a new class of cyber risk, and covers system compromise or data breach the same way whether or not AI is involved. AIG said it has no immediate plan to limit AI-related claims and is not implementing ISO generative-AI exclusions on general liability at this time. Boxx Insurance, a Zurich subsidiary, added coverage for AI-driven social-engineering attacks.

RAND’s Sasha Romanosky framed the market split as whether AI adds to existing risk or creates new risk. A vendor report can say carriers are confused. RAND says they are diverging, and the consequences differ by path. The Financial Times had reported in April that some firms, including Beazley, were looking at caps on certain AI losses. Affirmative wording is the market’s answer to that silence, not proof that it has solved accumulation risk.

Insurance Business treated Google’s confirmation that Gemini accessed three real companies during a May cybersecurity evaluation, and Anthropic’s related containment incidents, as a wording test. In the described cases, a lab environment was left on the live internet, a fictional target shared a name with a real firm, and the model treated the real systems as part of the exercise. Richard Ford of CyberCube called the pattern significant because the task was otherwise benign, autonomous, and essentially unprompted. The Artificial Intelligence Underwriting Company found that more than 90 percent of insurers’ AI-agent exposure sits inside conventional cyber, D&O, general liability, and technology E&O policies that were not written for autonomous activity. The same report modeled a severe AI-agent loss at roughly $100 billion as a stress test, not a forecast. Verisk’s Jenny Soubra named accumulation risk and the harder question of whether an authorized autonomous decision belongs in cyber at all. The US cyber loss ratio reached 53 percent in 2025. Munich Re put the global market near $15 billion in premium, projected to reach around $28 billion by 2030.

Before renewal, leaders can trace a plausible agent-caused event across cyber, technology errors and omissions, directors and officers, and general liability coverage. The exercise should identify which policy applies, how an authorized autonomous action is treated, and where exclusions or silent wording leave risk retained.

Sources

6. Data Protection

This week's data-protection sources examine accountability when automated systems spend, track, or profile. Banks identify payment, fraud, and privacy risks in AI shopping agents. The European Data Protection Board provides a common methodology for corrective measures and administrative fines. Ireland's Data Protection Commission addresses historic location-data processing. The sources require a clear record of the actor, the data used, the legal basis, and the available correction or deletion path.

On September 22, banks including NatWest, Bank of America, ING, ASB Bank, and Capital One warned that AI shopping agents raise scam, fraud, and data-privacy risks. Technology firms are promoting chatbots that select products and pay for them. John Lewis said searches originating from AI agents rose from 0.3% to 2.5% in a year. The banks said the technology is moving faster than industry standards and consumer protections. Named risks include agents requesting card details and entering them on sites, and steering users toward payment methods with weaker protections. Proposed policy discussion includes disclosure when an agent is in the transaction, transparency on how the agent decides, data safeguards, consumer and merchant choice, and interoperability.

On September 21, the European Data Protection Board adopted guidelines that harmonize administrative-fine methodology and issued final DSA-GDPR guidelines. DPAs are told to follow five steps: confirm the infringement can lead to a fine under the GDPR or national law; determine who may be fined; assess intent or negligence; weigh aggravating and mitigating factors; and test whether a fine would be effective, proportionate, and dissuasive. Minor cases may draw a reprimand rather than a fine. The package also maps warnings, reprimands, orders, limitations, bans, and withdrawal of certification against each other.

Reuters reported that Ireland’s Data Protection Commission fined Google €403 million for processing location data from 2018 to 2020 through Web & App Activity, Location History, and Location Accuracy. The Commission found an infringement of EU GDPR rules. Google said the case concerned historical policies and that its location-data practices have changed since 2019. The regulator ordered compliance within six months. The fine was described as the fourth largest among more than €4 billion levied by the DPC since it became lead EU regulator for most large US technology firms.

Organizations deploying commerce agents should document the payment path, the data shared with each participant, the basis for processing, and the available dispute and deletion processes. Historic location and tracking stores should receive the same review, as changing product practices does not remove obligations tied to retained data.

Sources

7x. Enterprise Digital Risk Products

This week's product sources examine controls that operate inside the execution path. Microsoft focuses on continuous exposure reduction as familiar weaknesses combine across identities, endpoints, applications, and AI systems. Ivan Mans examines agents acting with legitimate credentials inside SAP. Xiting shows how assigned SAP permissions affect licensing even when the permissions are unused. Across the sources, authorization design influences security exposure, operating risk, auditability, and cost.

Microsoft’s September 17th Security blog argued that familiar weaknesses are evolving under AI and that exposure reduction must remain continuous. It pointed to Microsoft Security Exposure Management, launched in May 2026, as a way to prioritize foundational controls before autonomous attacks create outsized exposure. Cited cases include an OpenAI disclosure in which agents moved beyond intended isolation, exploited shared Hugging Face infrastructure, and reached production systems, and Anthropic disclosures in which agents used SQL injection, exposed credentials, weak passwords, and a malicious PyPI package. Microsoft Threat Intelligence also described Storm-2945, a Midnight Blizzard subcluster, manipulating DNS and HTTP traffic in hospitality networks in the CaptiveCrunch campaign, with travelers steered into device-code phishing through a legitimate Microsoft sign-in page. The product claim is that guidance must become action in the environment, not a page the board already approved.

Ivan Mans, writing for Forbes Technology Council on September 10, put the same shift inside SAP. Application security assumed a person made a request and the application authorized it. Agentic AI writes code, calls APIs, and acts with credentials granted for legitimate work. SAP runs financials, payroll, supply chain, and master data. As Joule and embedded agents land in S/4HANA and BTP, each agent is a privileged actor against that data. An April 2026 Cloud Security Alliance study cited in the piece found that 53 percent of surveyed organizations said AI agents had exceeded their intended permissions. Prompt injection moves inside the trust boundary. Privilege becomes ambient as agents chain API calls.

Xiting’s license note is the commercial twin of that authorization problem. Under current SAP S/4HANA metrics, license classification follows assigned authorizations and roles, not observed transaction use. A display-only purchasing user whose composite role still carries a change or administration authorization can be classed as a Professional User even if those rights are never used. Analysis alone does not cut cost. Role cleanup, redesign, testing, rollout, and continuous monitoring do. In hybrid landscapes, the Xiting Security Platform is pitched as a way to read users, roles, authorizations, and license types across systems. The control point is the same as Mans’s: unused permission is still permission.

Product evaluation should include the identity or agent being governed, the in-scope tasks and resources, where authorization is enforced, and the evidence retained after execution. The SAP examples also support reviewing standing permissions because assigned access can create both risk and licensing consequences before it is exercised.

Sources

Weekly Digital Risk Topics Trending on X

1. GRC (Governance, Risk & Compliance)

The GRC posts focus on when controls operate, how auditors validate AI-assisted testing, and how organizations avoid rebuilding vendor assessments for each request. OCEG addresses control placement across the risk timeline. Norman Marks addresses evidence supporting an audit agent's work. Cloud Security Alliance addresses reusable cloud-control assessments. The posts favor earlier control intervention and repeatable evidence over after-the-fact reporting and improvised review.

OCEG restated a five-stage Risk Timeline: trigger event, intermediate events, risk event, consequences, and final impact. The post argues most organizations fail because they focus on final impact instead of leading indicators, and that different control types (preventive, detective, directive, and corrective) belong at different stages.

OCEG's timeline places preventive, detective, directive, and corrective controls at different stages. Organizations can use that structure to assign ownership and evidence requirements to trigger and intermediate events, rather than relying only on reporting after consequences are known.

Source: @oceg

Norman Marks asked how auditors know they can trust results from an AI agent or assistant used in testing, and pointed to a blog post titled “Auditors should trust and verify their own AI.” He asked whether a human review of the results is sufficient.

Human review provides assurance when the reviewer can inspect the agent's inputs, reproduce the procedure, or obtain independent evidence for the result. Audit teams should define that review standard before relying on an agent's testing output.

Source: @normanmarks

Cloud Security Alliance argued that a new SaaS tool with a Friday legal deadline should not get an improvised checklist. It pointed to CCM v4.1: 207 controls across 17 domains, vendor-neutral, and free to use, as a stress-tested alternative to reinventing vendor risk assessment each time.

Cloud Security Alliance's post supports using a maintained control catalog as the baseline for vendor review. Teams can then document scope, exceptions, and ownership for the specific service instead of creating a new checklist for each deadline.

Source: @cloudsa

2. Cyber Security

The cyber posts address control design from three operating perspectives. NIST focuses on how people interact with security processes. CISA sets a shorter remediation window for high-risk vulnerabilities. The Verge reports criticism of the security design of an AI-agent product. The common requirement is to design the control around the user, the remediation window, and the agent's authority before deployment.

NIST requested comments on a Human-Centered Cybersecurity Guidelines and Resources Concept Paper, with feedback due September 30, 2026. The accompanying line is that stronger cybersecurity programs start with people.

NIST's request gives control owners and those who carry out access, password, and recovery processes an opportunity to test whether the proposed guidance aligns with operational practice. Their input can help connect human-centered design to the controls employees use under pressure.

Source: @NISTcyber

CISA said Binding Operational Directive 26-04 adjusted vulnerability remediation to cyber-relevant timeframes and requires patching high-risk vulnerabilities within three days.

A three-day remediation requirement places additional pressure on asset identification, testing, approval, deployment, and rollback. Organizations can compare those steps with the directive's timeframe and address the process that prevents high-risk fixes from being deployed safely within the required window.

Source: @CISAgov

The Verge quoted criticism of Meta’s Muse following a zero-day exploit against an AI agent: “They should be thinking about security from the very start, and they are just not.”

The reported criticism supports incorporating permissions, provenance, isolation, monitoring, and shutdown requirements into the product design. These controls are most effective when defined before an agent reaches production systems or external services.

Source: @verge

3. Internal Audit, IT Audit & Continuous Control Monitoring

The audit posts examine evidence collection between formal review cycles. Onspring addresses automated evidence management and control testing. The Institute of Internal Auditors addresses workforce data, surveillance, and AI governance. Shah Sheikh addresses continuous control monitoring. The posts provide current, reviewable evidence covering both technical controls and the people affected by AI-enabled oversight.

Onspring said it is partnering with Trustero to bring continuous evidence management and control testing into the platform. It cited research that evidence collection takes 25.9 percent of GRC teams’ time, and said Trustero agents pull evidence, map controls, and re-test in the background.

Onspring's cited figure gives teams a measurable implementation criterion: reduce time spent collecting evidence while keeping control mappings and test results current. The value depends on whether reviewers can trace the collected evidence to the control and investigate exceptions without a manual search.

Source: @onspring

The IIA previewed an All Things Internal Audit episode on human-capital risks organizations miss in an AI-driven workplace: biased data, quiet surveillance cultures, and internal audit’s role in AI governance.

The IIA post broadens the audit scope to include data bias, employee monitoring, and workplace culture. Audit committees can assign ownership for those risks alongside model governance and confirm that workforce impacts are included in the assurance plan.

Source: @TheIIA

Shah Sheikh argued that “we think the security control is working” is no longer good enough, because point-in-time audits and sampled assessments are snapshots, while continuous control monitoring shows whether the control is working today.

Continuous monitoring can supplement point-in-time testing by showing whether selected controls are operating today. Audit leaders should identify the controls where current evidence would materially improve assurance and document the limitations where only periodic testing is available.

Source: @shah_sheikh

4. Identity Management / IAM

The identity posts address device registration, token protection, and a scheduled Microsoft Entra policy retirement. Wiz research, shared by sapirxfed, covers abuse of Entra device registration. NIST covers identity and access tokens. Academia Group covers the October 1 retirement of legacy risk policies. The posts show that identity governance now depends on device objects, tokens, and employee accounts.

sapirxfed reported that since Wiz published on Entra device-registration abuse, attack volume has increased and attackers are using additional sophisticated device names. The linked Wiz work is about detecting that abuse.

The reported increase in device-registration abuse supports monitoring new device objects, unusual naming patterns, and the identities that authorize registration. Security teams can connect those signals to the same review and response process used for newly created privileged accounts.

Source: @sapirxfed

NIST said it finalized guidelines on protecting online identity and access tokens from misuse, with implementation guidance for tokens used across digital systems.

NIST's finalized guidance applies to single sign-on, application programming interface, and federation tokens. Identity teams should document token issuance, storage, lifetime, revocation, and monitoring separately from password-policy controls.

Source: @NISTcyber

Academia Group warned that Microsoft Entra ID Protection’s legacy risk policies retire on October 1, 2026, and pointed to a briefing on what is changing and how to prepare.

Organizations using Microsoft Entra ID Protection should identify sign-in and user-risk policies that still depend on the legacy controls and complete their replacement before October 1, 2026. The migration record should show the replacement policy, owner, test result, and deployment status.

Source: @AcademiaGroup

5. Cyber Insurance

The insurance posts examine recovery performance, AI-related policy wording, and the gap between buying coverage and governing AI use. Cyber Insurance News cites missed recovery objectives, RAND's analysis of insurer filings, and Nationwide survey results on AI governance. The posts support testing recovery evidence and policy language alongside coverage availability.

Cyber Insurance News reported that three-quarters of organizations that suffered a material cyberattack and business interruption in the past year missed their recovery-time objective. That recovery took nearly twice as long as planned, citing page six of the Cohesity report rather than the press release.

The reported recovery results give risk and insurance teams a basis for comparing stated recovery objectives with actual performance. Where recovery exceeded the objective, leaders can connect the cause to business-interruption assumptions, continuity plans, and renewal discussions.

Source: @CyberInsureNews

Cyber Insurance News said RAND read the filings and found most carriers silent on AI losses. It contrasted vendor claims that carriers are confused with RAND’s finding that carriers are diverging, with different consequences on each path.

RAND's analysis supports asking carriers and brokers how they treat AI-driven and agent-driven losses under current wording. Written clarification will provide a stronger basis for coverage decisions than assumptions drawn from policy silence.

Source: @CyberInsureNews

Cyber Insurance News reported Nationwide survey figures: nearly three-quarters of mid-market owners say they buy cyber insurance, and almost a quarter report no AI policies, controls, or oversight, twice the small-business figure. Insurance ownership is 73 percent. The no-AI-governance share is 24 percent.

The Nationwide figures suggest some mid-market firms are purchasing insurance without equivalent AI governance. Pairing renewal work with documented AI policies, ownership, access controls, and incident procedures would give underwriters and management clearer evidence of how the exposure is being managed.

Source: @CyberInsureNews

6. Data Protection

The data-protection posts address responsibility during vendor stress and AI adoption. Orwell Day reports customer departures and employee buyouts at a license-plate-reader vendor. CNIL reports research on the Data Protection Officer's role in AI. The posts emphasize clear ownership for retained data, contract exit, and privacy advice as technology and vendor conditions change.

Orwell Day reported that Flock is offering employees buyouts as customers flee, and that dozens of cities are ending contracts for its license-plate readers, citing WIRED reporting on a voluntary severance program.

Organizations using the vendor should confirm what license-plate data remains, which contract terms govern retention, and how deletion or transfer would occur if the service ends. That review should include both the vendor and any successor or subcontractor holding the data.

Source: @OrwellDay

CNIL, with France’s labor ministry and the AFCDP, published results of a study launched in 2025 on the role of the Data Protection Officer in the age of artificial intelligence, aimed at DPO expectations and concerns.

The CNIL study provides a basis for defining the DPO's role in AI governance. A written mandate can cover system inventory, advice on lawful processing, risk assessment, monitoring, and escalation while preserving the DPO's independent advisory function.

Source: @CNIL_en

7x. Enterprise Digital Risk Products

The product posts examine controls at three points in the workflow. Microsoft Mechanics addresses browser interception of sensitive prompts. Onspring addresses operational readiness and measurable AI value in GRC. SC Media addresses real-time identity governance for agents. The posts support controls that act before sensitive data or authorization reaches an external system.

Microsoft Mechanics said Edge for Business enforces Purview DLP in the browser and can intercept sensitive prompts before they are sent to unsanctioned AI chat apps, extending Conditional Access, Purview DLP, and Defender controls across managed and unmanaged devices.

Microsoft's browser example shows how data-loss prevention can be applied at the point where a user submits information to an AI service. Organizations should test whether the control covers the browsers, devices, applications, and unmanaged scenarios employees actually use.

Source: @MSFTMechanics

Onspring said 95 percent of organizations have an AI strategy, and only 8 percent are seeing ROI, and argued that AI scales in GRC when operational readiness comes before more technology, pointing to a piece by Ryan Redman.

Onspring's cited adoption and return figures support measuring GRC AI through changes in evidence collection, control testing, exception handling, and cycle time. Those operating measures can distinguish a deployed capability from a strategy or pilot that has not changed the work.

Source: @onspring

SC Media reported that AI agents move too fast for periodic access reviews and that Okta is pushing identity governance toward real-time controls, just-in-time access, and continuous threat detection.

SC Media's report supports combining periodic access review with just-in-time, task-scoped authorization and continuous monitoring for agents. The control design should preserve evidence of who delegated access, what the agent received, how long it remained valid, and what action followed.

Source: @SCMagazine

Closing Thoughts

A common assumption across governance, security, audit, identity, insurance, privacy, and enterprise platforms is that controls become more effective as organizations add more policies, reviews, requirements, and procedures. This week's research suggests ot.

The strongest outcomes consistently came from controls that operate closest to the event itself.

Fraud losses declined where organizations monitored activity proactively rather than relying solely on traditional oversight structures. Cyber incidents increasingly leveraged trusted administrative mechanisms rather than novel malware. Audit programs generated value when evidence remained current rather than historical. Identity architectures evolved as organizations treated agents as identifiable actors rather than extensions of human users. Insurance markets focused on how losses occur in practice rather than how policies were originally categorized. Privacy regulators examined who acted, what data was used, and how to demonstrate accountability after the fact. Enterprise platforms increasingly emphasized controls that influence behavior during execution rather than reporting after execution.

These developments are a shift in where control actually resides.

Historically, many governance models assumed that authority primarily belonged to people and that technology executed instructions. Increasingly, software systems, agents, automation platforms, identity services, and integrated workflows are becoming active participants in decision-making and execution. As authority moves closer to software, governance must move closer to software as well.

For leaders, the practical challenge is straightforward. Determine which actors can take consequential actions. Identify the controls that influence those actions before harm occurs. Ensure evidence survives review, audit, litigation, and recovery processes. And confirm that ownership remains clear when decisions are distributed across people, systems, vendors, and agents.