Highlights: Proving controls while AI amplifies risk

Executive Summary

Digital risk management is entering a new phase. Generative AI is no longer an emerging technology sitting at the edge of the enterprise. It is becoming embedded in business processes, security operations, audit programs, identity systems, claims platforms, and core enterprise applications. At the same time, threat actors are adopting the same technologies to automate reconnaissance, refine attacks, scale fraud, and accelerate compromise. The result is not a fundamentally new risk landscape, but a rapidly amplifying one in which existing vulnerabilities, governance gaps, and identity weaknesses can have greater operational and financial consequences than ever before.

This report examines the most significant developments across governance, cybersecurity, audit, identity, insurance, data protection, and enterprise risk technologies. While the topics vary, a common challenge emerges throughout: organizations must move beyond compliance, policy creation, and technology deployment to demonstrate that controls remain effective in practice, residual risks are understood, and resilience can be evidenced in real time. As AI adoption accelerates and regulatory, operational, and threat pressures converge, the ability to prove trust, accountability, and control effectiveness is becoming a defining measure of organizational maturity.

Deep Research

1. GRC (Governance, Risk & Compliance)

KPMG International’s 2026 Global Chief Ethics and Compliance Officer Survey of 725 CCOs across six industries and eight countries describes a compliance function under compounding pressure from AI, misaligned regulators, and international uncertainty. Three themes dominate. First, operational durability is the organizing priority for interconnected cyber, third-party, and regulatory risk. 75% of CCOs are looking to cybersecurity and data privacy for additional investment; 77% cite data analytics. In the United States, those figures rise to 87% and 85%. Second, CCOs are cautiously optimistic about AI: 68% feel “mixed, leaning positive,” and another 24% report considerable efficiency gains, yet use remains concentrated in lower-sensitivity work such as risk assessment (50%), visualization (44%), and training (44%). Hotlines and investigations sit at 4%. Third, 67% of CCOs are confident they can assess compliance synergies across legal, human resources, investigations, internal audit, and operations, continuing the shift from cost center to value driver. New regulatory requirements are the top two-year challenge (33%), followed by data analytics and predictive modeling (31%).

That investment story still needs a language the board will fund. FAIR Institute Standards Director Todd Tucker argues that vulnerability management’s gap is not tooling. It is translation. Security speaks in CVSS scores and exploitability; the business speaks in dollars, priorities, and trade-offs. FAIR’s primary value is not Monte Carlo simulation. It is a shared taxonomy that separates loss-event frequency from loss magnitude and lets a CISO frame a few material scenarios instead of thousands of “critical” findings. Quantification should be reserved for exceptions—systemic exposures or control decisions that add customer friction—while routine patching stays fast and automatic.

NIST’s initial public draft of SP 1353, Quick-Start Guide for Using Artificial Intelligence (AI) for CSF Analysis and Reporting (August 2026), shows how generative AI can support Cybersecurity Framework 2.0 current-state and target-state profiles using CO-STAR prompts. The draft makes it clear that the examples are notional, that AI output is not an assessment methodology, and that qualified personnel must review generated content before it informs decisions. For GRC leaders, the operating implication is consistent: invest in resilience and analytics, speak risk in business terms, and treat AI as an assistive analyst under human accountability rather than an unattended control owner.

Sources

KPMG 2026 Global CCO Survey

2. Cyber Security

Island’s July 2026 research describes a supply-chain shift that matters as much as any single CVE. Investigators found about 7,600 malicious GitHub repositories created by roughly 6,600 profiles. More than 800 posed as AI Skills or MCP servers and appeared more than 600 times in public AI registries. The wider FakeGit operation recorded more than 14 million measured downloads. The payload is familiar: SmartLoader establishes persistence and installs StealC to harvest credentials and sessions. The new technique is AgentBaiting. An AI agent searching for a Skill or MCP server can discover a campaign repository on its own, treat the attacker’s README as documentation, and hand installation instructions to the user. In testing, Claude Code, Gemini, and ChatGPT all surfaced malicious repositories without being shown a link. A playbook built to deceive people now deceives the agents acting on their behalf.

That pattern is already visible in live crime. Cisco Talos disclosed UAT-10147, a previously undocumented Chinese-speaking, financially motivated group active since early 2026. It exploits known flaws in internet-facing systems, including Zimbra, Nacos, Telerik UI, and AjaxPro; deploys a cross-platform implant called SPECTRE plus NoodleRAT, QuasarRAT, and related tooling; and uses agentic AI for reconnaissance, exploit refinement, payload generation, troubleshooting, and persistence. On Windows, SPECTRE can blind EDR; on Linux it loads a persistent kernel rootkit. Researchers recovered a target list of about 170,000 URLs spanning government, education, media, technology, and gaming organizations.

Policy and market structure are moving at the same time. An August 12th Presidential Memorandum authorizes vetted private firms, under federal contract and supervision, to conduct cyber surveillance and cyber-effects operations against cyber-enabled transnational criminal organizations. The White House fact sheet cited more than $20.8 billion in reported U.S. losses to cyber-enabled crime in 2025 and noted that 73% of U.S. adults report experiencing an online scam or attack. The memorandum creates legal risk no company has been asked to carry, and no court has tested, because the Computer Fraud and Abuse Act generally bars private actors from touching attacker systems. Separately, Techstrong’s Black Hat 2026 special report asks whether the industry is too big for its own commercial machinery: more than 4,100 vendors offer about 11,000 products, while 2025 vendor funding reached $13.97 billion across 393 rounds and first-half 2026 added $7.41 billion. Size is not the same as progress. For security leaders, the practical response is tighter control of the AI-capability supply chain, faster checking against known exploited software, and clearer governance before any firm considers participating in government-directed offensive work.

Sources

3. Internal Audit, IT Audit & Continuous Control Monitoring

KPMG Belgium’s guidance is blunt: high-impact AI must be demonstrably trustworthy, and policy reviews are no longer enough. Traditional audit methods assumed stable systems and clear control points. Generative and agentic AI rely on dynamic data, produce probabilistic outputs, and can orchestrate tasks across systems. Shadow AI, vendor-embedded features, and custom models each carry different risk. Internal Audit’s role is to move the organization from confidence based on frameworks towards evidence that systems behave as intended.

A mature plan should not rely on a single audit type. KPMG describes three. An AI governance audit asks whether the organization is structurally ready: vision, roles, inventory, policies, risk management, data, security, privacy, documentation, monitoring, and literacy. Common findings include unclear ownership, incomplete inventories, weak approvals, and poor visibility over third-party and embedded AI. An EU AI Act compliance audit starts with that same inventory, then classifies risk and tests documentation, human oversight, incident handling, and literacy against applicable obligations. Treated as a standalone legal checklist, it misses operational risk; embedded in a trusted-AI control framework, it connects regulation to evidence. An AI solution audit asks whether a specific system is designed, implemented, used, and monitored in a controlled way, combining qualitative governance review with quantitative tests such as fairness checks, class-imbalance analysis, and robustness testing. LLM audits add hallucination, prompt manipulation, and leakage. Agentic audits use KPMG’s TACO framing—Taskers, Automators, Collaborators, Orchestrators—and emphasize unique identifiers, immutable logging, and kill-switch protocols.

Picus’s Blue Report 2026, based on more than 338 million attack simulations from January to June 2026, supplies the control-effectiveness evidence audit committees should demand. Average prevention recovered from 62% to 69%, returning to its 2024 peak. Post-compromise prevention was only 37%. Stealth prevention covering discovery and collection was 10%. Endpoint prevention rose to 83%, while malware-download prevention fell to 50%, and North America posted the lowest regional prevention score. Every one of the ten least-prevented ransomware families scored 38% or lower; Play collapsed from 50% to 13%. The report’s lesson matches KPMG’s: controls degrade without continuous validation. For internal audit, that means pairing governance audits with live evidence that prevention still works after the attacker is inside.

Sources

Picus Blue Report 2026

4. Identity Management / IAM

Identity programs were built for two columns: humans with a joiner-mover-leaver lifecycle, and machines with a service account, a purpose, and an owner. AI agents sit in neither. A CIO.com analysis published 21 August 2026 notes that assigning an agent the invoking user’s permissions works only for short tasks and breaks as agents run longer or touch more systems. A shared service account solves delegation but creates standing privilege that outlives the task. The more mature target is a workload identity that is short-lived, tightly scoped, and ephemeral. Even that model assumes foreseeable behavior. Agents do not. Actual access can shift mid-task depending on the prompt, the tool chosen, or a plugin. Parent agents can hand slices of permission to child agents until the resulting chain exceeds what any approver contemplated. Non-human identities already outnumber humans; Cloudflare has reported that automated traffic has overtaken human requests on its network.

Cisco Duo’s 2026 CISO Perspectives report, based on interviews with more than two dozen CISOs, treats identity as both the problem and the change agent. Identity security is a top-three priority for most organizations. Leaders flagged the traditional joiner-mover-leaver lifecycle, third-party access, and non-human identities as drivers of sprawl; more than half expressed concern about managing leavers. A prior Duo study found 86% of leaders concerned about inadequate contractor and third-party controls. Every CISO interviewed mentioned agentic AI. One head of identity governance said it is “not two or three years away; it’s here, and there’s little governance over it.” Tokens are not sufficient. Unless security sits between the agent and the resource, nothing can be enforced or observed in real time. Legacy Active Directory debt compounds the issue: decades of borrowed service accounts become dangerous when every new agent reuses them.

The week’s tactical signals reinforce the same foundations. DCSync remains a standard ransomware and nation-state move because replication rights, not only NTDS.dit protection, determine whether hashes leave the domain. CVE-2026-18963 shows that even modern identity platforms can expose unauthenticated account takeover through a password-reset flow. WhatsApp’s multiple-passkey rollout is a consumer-scale example of phishing-resistant authentication moving into daily use. IAM leaders need both: clean up standing privilege and AD baggage, and extend Zero Trust—zero standing privileges and continuous verification—to agents that can request their own tokens at machine speed.

Sources

5. Cyber Insurance

Claims operations remain the cost and customer-experience core of property and casualty insurance. Liberate’s guide on AI claims agents describes why the traditional model is under strain. Loss-adjustment expenses consume 25 to 30% of earned premiums in some lines. The U.S. P&C industry ran a 25.2% expense ratio in 2024. Roughly half the P&C workforce is expected to retire by 2028. Claim volumes spike during catastrophes while staffing stays fixed. PropertyCasualty360 has estimated that poor claims experiences could put up to $170 billion in annual premiums at risk over five years. AI claims agents differ from chatbots and interactive voice response. They complete work such as first notice of loss intake, coverage verification, structured data capture, triage, vendor dispatch, and status updates across voice, SMS, email, and web, with context persisted across channels. Branch Insurance reported a 42% reduction in call-resolution time and 70% cost savings after deploying Voice AI and digital FNOL with Liberate. Allied Trust completed digital FNOL in six weeks. Governance is not optional. Responsible platforms require authority limits, full transcripts and audit logs, human-in-the-loop escalation, and certifications such as SOC 2 and PCI.

The same week’s loss activity shows why underwriting still depends on identity hygiene. Apollo Global Management disclosed unauthorized access to certain cloud platforms between 6 and 10 July. Potentially affected data included names, dates of birth, contact information, home addresses, and Social Security numbers. Apollo was among dozens of financial firms targeted by ransom-seeking actors using phone calls and lookalike password-harvesting sites. Affected individuals were offered identity-protection and credit-monitoring services. Uber Freight and Levi Strauss separately reported unauthorized-access investigations. Low-tech social engineering remains effective alongside AI-driven threats.

Clyde & Co’s Corporate Risk Radar, covering 700 senior decision-makers across 11 industries, quantifies how fast the risk conversation has moved. Eighty-six% of leaders now rate technological risk as “high impact,” up from 46% a year earlier—the largest year-on-year increase of any category. Sixty-eight% call transformational delivery a high risk; 57% feel pressure to prove return on technology investment; 72% cite technology implementation and systems integration among their greatest operational challenges. Geopolitical risk is no longer background monitoring: 72% say it is having a direct commercial impact, versus about 50% in 2025. Partner Tim Crockford noted that an AI governance framework that looks mature today can be out of date tomorrow. For brokers, carriers, and insureds, the implication is dual: AI can compress claims cost and cycle time if it is auditable, and the same technology, plus third-party concentration and voice-based intrusion, is now a board-level underwriting issue.

Sources

6. Data Protection

Two disclosures this period illustrate how customer, employee, and healthcare data remain key targets even when negotiations fail or the full scope is still unknown. ShinyHunters claimed it stole 50 gigabytes of Carhartt data after the workwear company rejected further talks on a $3.3 million demand. The alleged haul included customer and employee personal information, loyalty data, and internal corporate information. Have I Been Pwned later analyzed the leaked dataset and identified 12,933,413 unique email addresses likely belonging to real users—about 52% of the file. Founder Troy Hunt excluded 11,942,664 synthetic, test, disposable, and non-human addresses, nearly 48% of the original data set. The split is operationally important. Breach-notification and monitoring programs should focus on the verified human population. Still, the presence of millions of non-human addresses also shows how noisy extortion dumps can inflate headline counts if they are not filtered.

Nutex Health, a for-profit operator of 28 facilities across 12 states with $875 million in 2025 revenue, disclosed a cyberattack in a U.S. Securities and Exchange Commission filing dated around August 24th. Preliminary results show that an unauthorized third-party accessed and exfiltrated information from company servers, including some information that may be private or confidential. Nutex hired external incident-response and forensic specialists, activated its response plan, implemented containment, and notified law enforcement. The company has not yet determined whether patients, employees, credentialed providers, confidential business and financial information, or intellectual property were involved. As of August 24th, it reported no material impact on operations or financial reporting systems and said it does not currently believe the incident will materially affect strategy, operations, or results. BleepingComputer found no threat actor publicly claiming the attack.

Together the cases argue for differentiated data protection rather than a single “encrypt and notify” playbook. Retail loyalty and customer email files create long-lived identity-fraud exposure once published. Healthcare operators face unresolved scope questions that can trigger HIPAA, state, and securities obligations before the data types are fully classified. Continuous inventory of what actually sits on production servers—and the ability to separate real human identifiers from synthetic noise—determines both regulatory response and the quality of evidence provided to insurers, auditors, and customers.

Sources

7x. Enterprise Digital Risk Products

Onapsis’s State of AI, Security and ERP pulse report captures a market moving faster than its own confidence. More than half of respondents began using AI applications or agents that touch ERP within the last six months; another 41% are directed to do so before year-end. Efficiency is the stated driver (78.3%). 86% have integrated or will shortly integrate AI directly into ERP code; 62.3% already use AI-generated code in those systems. 75% know agentic workflows are touching ERP because they put them there; 15% think it is possible but cannot confirm. Trust is being placed outside the enterprise: 83.3% plan to use native vendor code and 60.6% will source third-party applications. 41% of cybersecurity teams are slowing AI-in-ERP adoption because they lack confidence in security and compliance controls. Nearly 57% report at least one business unit objecting, led by security (41.4%) and IT (20.7%), citing lack of confidence in AI security (75%) and compliance risk (71.6%). 21.6% already report confirmed incidents in which attackers used AI against business-critical platforms. Yet 68.6% of cybersecurity leaders are only “somewhat” or “not very” confident their defenses could detect an AI-based attack, and more than 70% have only some or no trust in AI to secure critical data. At the same time, 84.4% call deploying AI to help secure ERP a top 2026 priority.

That confidence gap collides with an approaching identity deadline in SAP landscapes. Xiting’s 24 August guide states that SAP Single Sign-On 3.0 reaches end of maintenance on 31 December 2027, including the NetWeaver AS Java Secure Login Server. SAP Secure Login Service is the cloud-based successor. Kerberos and SPNEGO are not Zero Trust capable: they rely on offline tickets and lack native multi-factor authentication, device trust, and conditional access. SAP GUI’s SNC protocol does not support SAML or OpenID Connect, a gap SLS is designed to close by issuing short-lived, non-exportable X.509 certificates after identity-provider, MFA, and device checks. Xiting notes that stolen or weak credentials cause more than 80% of security incidents and that forgotten SAP password resets cost more than 11 hours of productivity per employee per year. A planned coexistence phase between SSO 3.0 and SLS is the practical migration path.

For enterprise risk-product buyers, the combined message is simple. AI is already inside ERP code and agent workflows, vendor trust is high, detection confidence is not, and core SAP authentication is on a known countdown. Programs that treat evidence, identity modernization, and AI-ERP security as one operating model—not three projects—will be in a stronger position before the 2027 deadline and before the next agent-driven incident.

Sources

SAP Data Masking and Anonymization

Weekly Digital Risk Topics Trending on X

GRC discussions this week mixed residual-risk discipline with the practical reality that enterprise generative AI amplifies existing ransomware paths rather than inventing new ones. OCEG reminded practitioners that residual risk—the exposure that remains after controls are working—is what executives actually lose sleep over, and distinguished inherent risk, control risk, detection risk, and residual risk as the core vocabulary of professional GRC work (OCEG on residual risk). BleepingComputer, citing Acronis, argued that identity controls, governance, and least-privilege access are the containment strategy for AI-enabled ransomware rather than treating AI itself as a novel threat class (BleepingComputer on AI and ransomware). The ACFE reported that Operation Jackal IV produced 58 arrests and identified 263 suspects in an eight-month effort against West African organized-crime money-laundering networks, an indication that fraud and financial-crime risk remain part of the same compliance landscape (ACFE on Operation Jackal IV).

Cybersecurity posts focused on operational guidance, leadership, and a new financially motivated actor using AI in live attacks. NIST published tips and tactics for building-automation and control-system cybersecurity (NIST building-automation guidance). Dark Reading’s Heard It From a CISO series featured Standard Chartered Group CISO Cezary Piekarski on the shift from technical practitioner toward strategic security leader (Dark Reading CISO interview). Daily Dark Web summarized Cisco Talos’s disclosure of UAT-10147, a previously undocumented Chinese-speaking cybercrime group active since early 2026, targeting Windows and Linux web servers internationally, recovering a target list of about 170,000 URLs, and integrating agentic AI into reconnaissance, exploit refinement, payload generation, and persistence (Daily Dark Web on UAT-10147). BlackRoomSec flagged CISA’s 21 August update to internet-exposure guidance for IoT, SCADA, ICS, and OT, plus the 25 August advisory “A Tale of Two SOCs” drawn from two red-team assessments (CISA updates via BlackRoomSec).

Internal audit and continuous control monitoring conversations emphasized testable AI controls and reusable assessment frameworks. Dan Kornas highlighted OWASP AISVS, a community catalog of 191 verifiable security requirements across 12 chapters and three assurance levels, covering data integrity, model supply chains, agentic action, and MCP security (OWASP AISVS). Cloud Security Alliance pointed teams to CCM v4.1—207 vendor-neutral controls across 17 domains—as a way to stop reinventing security questionnaires (CSA CCM v4.1). The IIA published Nancy Hom’s three-step checklist for deciding when it is appropriate to let AI take the wheel and when human review must remain in control (IIA on responsible AI use).

Identity and access management posts returned to classic Active Directory tradecraft and newer authentication flaws. Cyber Security News explained DCSync (MITRE ATT&CK T1003.006): an attacker impersonates a domain controller over MS-DRSR and walks away with NTLM hashes, Kerberos keys, and password history without touching NTDS.dit on disk (DCSync explained). ThreatWire flagged CVE-2026-18963, a CVSS 9.1 unauthenticated Keycloak account-takeover issue in the password-reset flow, with fixes in 26.7.2, 26.6.6, and 26.4.15 LTS (Keycloak CVE-2026-18963). Sami Laiho noted WhatsApp’s addition of multiple passkeys for phishing-resistant sign-in on iOS and Android (WhatsApp passkeys).

Cyber insurance discussion linked catastrophic operational loss to the current “Model T era of AI.” SquareX highlighted Martin Hinton’s conversation on the Marks & Spencer hack that cost £300 million and shut down operations, shadow AI, deepfakes, and North Korean fake IT workers (SquareX / Cyber Insurance News podcast). Force4 Technology Communications pointed to the “Insuring Catastrophic Cyber Risk” report and its five insights concerning federal action (Insuring Catastrophic Cyber Risk).

Data protection activity mixed a large alleged healthcare listing with rising privacy-compliance pressure. Daily Dark Web reported a threat actor advertising an alleged 1.3 million-user healthcare database for $2,000, with a schema consistent with CRM and Salesforce Health Cloud fields (Alleged 1.3 million healthcare records). Dr. Philippe Vynckier circulated TechTarget’s analysis that GDPR and AI are intensifying privacy and data-protection compliance obligations (GDPR and AI privacy demands). Nicolas Krassas highlighted 282-plus structured privacy and data-protection skills for AI agents following the agentskills.io open standard (Privacy skills for AI agents).

Enterprise digital risk product conversations focused on continuous decision intelligence and AI-delivery governance. Cyber Defense Magazine introduced LISN from Level 6 Cybersecurity as a platform intended to keep strategy, risk, and changing conditions connected between quarterly assessments (LISN continuous decision intelligence). OWASP thanked Chainloop for becoming a corporate supporter, positioning signed evidence, policy-as-code, and AI-assisted risk assessment as a way to ship AI-generated software faster and still prove it is secure (OWASP / Chainloop).

Closing

Across every discipline examined in this report, from GRC and internal audit to cybersecurity, identity management, insurance, data protection, and enterprise applications, the signals are clear. Organizations are being asked to make decisions faster, adopt AI more broadly, and operate in an environment where both innovation and threat activity are accelerating simultaneously. Success will depend less on the number of controls in place and more on the ability to demonstrate that those controls work when challenged.

The leading organizations will be those that can translate technical risk into business impact, continuously validate control effectiveness, manage human and non-human identities with equal rigor, and produce evidence of resilience as events occur rather than after the fact.

Want to discuss Digital Risk or AI topics in more detail for your organizations? Book a time that works for you: Schedule Call and Follow Me

#DigitalRisk #CyberRisk #GRC #ContinuousCompliance #CyberInsurance #IdentitySecurity #DataProtection #AIGovernance