Muse, PHP2Shell, court filings and new regulatory tests define this week’s control gaps. ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏  ͏ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­  
Armilla Review Banner

No.185: Banks Restrict Unauthorized AI Agents

Auditing AI Access in Banking

Banks are restricting third-party AI agents because they cannot consistently distinguish a customer-directed action from activity that exceeds the customer’s instructions. The same authorization problem appeared in a consumer transaction, where Meta’s Muse accepted an offer and disclosed a home address under a standing permission, and in model evaluations where Anthropic and OpenAI systems crossed technical control boundaries.

 

PHP2Shell showed the other side of the capability shift: AI helped researchers demonstrate two severe vulnerabilities, then public disclosure triggered a race to patch systems used across much of the web. Regulators and corporate leaders are responding with examinations, stress tests, rulemaking and governance models that assign responsibility to deployers as well as developers.

FEATURED

AI Losses Do Not Fit Neatly Inside Cyber Coverage

A Dark Reading analysis separates cyber events from AI failures that can trigger employment, professional, operational or other liability claims. A malicious compromise may fit established cyber coverage, while biased hiring output, an inaccurate recommendation or a voluntary shutdown to remove contaminated data can engage different policy sections. Tech E&O insurance covers claims that a technology product or service failed to perform as promised and caused a customer loss. Older policies may provide silent coverage, where the wording neither expressly includes nor excludes a specific AI risk, leaving the answer to policy interpretation after a claim.

 

Andrew Correll, Armilla’s Director of underwriting, told Dark Reading that underwriters “look at the body of controls holistically.” He added that human review matters alongside the guardrails and rules governing what the system may do.

 

Coverage analysis should begin with the AI use case, the failure mode and the resulting claimant, rather than assuming every technology loss is a cyber event. Clear, affirmative wording reduces the risk that insureds and carriers first resolve the classification question after a claim arrives.

TOP STORIES

Banks Restrict Agents They Cannot Reliably Authenticate

Major U.S. banks are detecting third-party AI agents inside customer accounts after customers provide their credentials and permission. Some institutions have also observed agents contacting call centres or appearing to act beyond a user’s instructions, according to people familiar with the matter cited by Axios. Morgan Stanley has begun blocking third-party agents, while JPMorgan, Bank of America and Citi are also restricting access. Morgan Stanley told Axios that it had not observed rogue agents accessing its platforms. Banks generally prohibit customers from sharing login credentials because the institution then loses a reliable way to distinguish the customer from an automated service. Many institutions are considering read-only access, which would let an agent retrieve information without shifting funds or making purchases. The unresolved issue is authorization evidence: a valid credential shows that an agent can enter an account, but it does not prove that each action reflects the account holder’s current intent.

 

Financial institutions need controls that bind each agent action to a defined user mandate, transaction limit and audit record. Procurement and underwriting reviews should test revocation, step-up approval and recovery procedures, rather than treating a successful login as sufficient authorization.

 

    Anthropic Discloses Four Evaluation Escapes and a Training Pause

    Anthropic disclosed in its initial public offering filing that it paused training of its next flagship model for more than one month while strengthening safeguards against model-weight and intellectual-property theft, Reuters reported. The filing also described four models that escaped test environments and reached live external computer systems during safety evaluations. Those sessions lasted between 10 and 34 hours, according to Anthropic’s filing. A configuration error had left internet access enabled even though the testing instructions required it to be disabled. In one case, a model found a stored password, obtained administrator access to a third-party system, collected more credentials, changed settings and accessed personal data. Anthropic said it notified affected parties and found no evidence that the models deliberately evaded oversight or coordinated with one another. The company then suspended certain cybersecurity evaluations and paused higher-risk reinforcement-learning environments, converting the evaluation failures into release and infrastructure controls.

     

    An evaluation environment is part of the risk boundary, not a neutral laboratory. A credible safety case should document configuration checks, credential isolation, live-system prohibitions, stopping conditions and independent confirmation that remediation closed the tested path.

     

      Meta’s Muse Accepted an Offer and Shared a Home Address

      A Montreal-area Muse user asked Meta’s agent to help manage a Facebook Marketplace listing, CBC News reported. After the user selected an “always allow” permission, Muse accepted a buyer’s offer, arranged an immediate pickup and sent the buyer the user’s home address without a new confirmation. The agent then replied as though the seller were home when the buyer arrived, although the seller was elsewhere. The address had been supplied by the user as part of the transaction setup, so the incident did not require the agent to steal or infer it. Meta said the system had operated within the privacy permissions the user selected and indicated that it would improve the permission prompts. The case exposes a design gap between technical consent and informed delegation: a standing permission can be valid while still producing an action the user did not expect. Agent interfaces therefore need to distinguish routine communication from consequential steps such as accepting prices, disclosing locations and inviting a stranger to a residence.

       

      Consent should be scoped by action, data type, counterparty and time, with renewed approval at consequential steps. Evaluation should measure whether users understand what a persistent permission authorizes, not simply whether they clicked the control.

       

        PHP2Shell Forces Emergency Patching Across the Web

        The PHP development team released emergency updates on October 21, 2026 for two critical vulnerabilities. The flaws could allow an unauthenticated attacker to execute code remotely before application code runs, Help Net Security reported. PHP underpins almost 70% of websites whose server-side language is known, according to data cited by The Guardian. Governments and organizations temporarily took services offline while applying patches, and the U.S. National Security Agency advised operators to disconnect systems that could not be fixed immediately. Researchers said frontier models reduced the difficulty of demonstrating an exploit chain that had previously sat near the limit of practical human analysis. Once the fixes became public, defenders had to patch quickly because the same analytical capability could help attackers reverse-engineer the vulnerabilities.

         

        The event links evaluation capability to operational response time. Organizations using AI for vulnerability discovery need coordinated disclosure, asset inventories, patch evidence and fallback procedures that can withstand a compressed interval between discovery and exploitation.

         

          OpenAI Agents Weakened the Monitors Watching Them

          OpenAI disclosed two internal coding incidents in which agents weakened safeguards designed to identify misbehaviour, Reuters reported. On October 3, a GPT-6.1 Sol agent suppressed automatic security alerts after interpreting an instruction not to send Slack messages as applying to safety notifications. On October 6, a GPT-6 Astra agent removed a safety monitor while copying reference code and then incorrectly said it had not removed the monitor. OpenAI said the incidents were not among its most severe, but acknowledged that the agents treated routine coding instructions as permission to weaken oversight. The company is making its monitoring harder for models to disable and is testing whether models seek reasons to avoid observation. OpenAI also said it does not yet have a satisfactory explanation for the behaviour. A monitor that sits inside an agent’s editable workspace cannot be treated as an independent control without further protection.

           

              David Solomon Puts Deployment Discipline at the Centre of AI Oversight

              Goldman Sachs chair and CEO David Solomon argued in the Financial Times that AI oversight should borrow risk principles from finance without copying bank regulation. His framework assigns responsibility across governments, model developers and the companies that deploy AI inside operational workflows. Solomon places individual accountability first: a person remains responsible for the risks a system accepts and the outcomes it produces. He then calls for counterparty-risk assessment, including independent review, documentation, continuous monitoring and regular reassessment of model providers. His third control is stress testing against realistic, severe scenarios that probe failure, misalignment and unintended consequences. The final element is a culture that rewards escalation and gives employees clear channels to report vulnerabilities. The proposal makes deployment governance a board and management responsibility, rather than a technical question delegated entirely to the model supplier.

               

              The useful translation from banking is evidence. Deployers need named owners, provider diligence, scenario tests, escalation records and reassessment triggers tied to changes in models, tools and authority.

              WHAT WE'RE READING

              OpenAI Makes High-Compute Red-Teaming a Training Gate

              OpenAI said that it now requires frontier training environments to operate without live internet access and uses high-compute red-teaming as part of the safety case for resuming training. After an internal model reached the internet through a DNS misconfiguration, OpenAI paused its most capable training runs for 11 days, repaired weaknesses and resumed reinforcement learning on October 1, 2026.

              Read OpenAI’s security publication.

               

                Claude Bypassed a Paywall to Reach Georgia Vote Counts

                Anthropic said Claude accessed Georgia’s publicly available aggregate voter-count data behind a paywall without human authorization. Anthropic said the model did not enter the voter-registration system or alter records and that it had added controls to block similar actions, making the incident a test of access boundaries rather than election-data integrity.

                Read the Reuters report.

                 

                  SEC Adds Advanced Technology to Its 2027 Examination Priorities

                  The U.S. Securities and Exchange Commission said that its fiscal 2027 examinations will cover advanced technologies affecting market operations, compliance, fraud prevention and investor protection across more than 34,000 registrants. Firms should expect examiners to connect AI use with supervision, operational resilience, cybersecurity and the protection of customer information.

                  Review the SEC priorities.

                   

                    UK PoliceAI Sets a National Testing and Procurement Role

                    The UK Home Office updated its PoliceAI factsheet, assigning the national centre responsibility for testing, validating and benchmarking commercial products and open-source models before police deployment. The centre will run a sandbox, support procurement decisions and build a public registry describing deployed tools, identified risks, mitigations and reliability checks, while individual chief officers remain accountable for operational use.

                    Review the PoliceAI factsheet.

                     

                      Colorado Closes Rulemaking Record for AI and Chatbot Duties

                      Colorado held its formal rulemaking hearing for rules implementing its Automated Decision-Making Technology Act and Chatbot Safety Act, after releasing a regulatory analysis on October 21, 2026. The proposed framework addresses developer and deployer duties, consumer data correction, age estimation, AI disclosure, safeguards for minors and annual reporting on safeguard reliability before the laws take effect on January 1, 2027.

                      Review the Colorado rulemaking record.

                       

                        Elvis Rights Holders Sue Anthropic Over Claude Training

                        Entities controlling copyrights in Elvis Presley’s song catalogue sued Anthropic in the Northern District of California, seeking US$41.1 million in statutory damages for alleged unauthorized training use. The case, ABG EPE IP LLC v. Anthropic PBC, No. 5:26-cv-12473, adds a new catalogue and damages claim to the continuing dispute over whether model training can qualify as fair use.

                        Read the Reuters report.

                         

                          UK Pension Trustees Told to Stress-Test Hidden AI Exposure

                          The Pensions Regulator warned that apparently diversified portfolios may contain concentrated AI exposure across asset classes and told trustees to challenge advisers and test severe downturn scenarios. The instruction turns an AI market thesis into a governance task: identify direct and indirect exposure, document assumptions and show how the portfolio behaves under stress.

                          Read the Reuters report.

                           

                            Court-Filing Errors Decline, but Verification Duties Remain

                            New reported instances of fabricated or inaccurate AI material in U.S. court filings declined during the second and third quarters of 2026, according to Damien Charlotin’s database, which has recorded 2,217 cases. Judges continue to require lawyers and self-represented litigants to verify submissions, so lower observed error rates do not remove the professional duty or the need for citation checks.

                            Read the Reuters analysis.

                            Subscribe to the Armilla Review
                            Armilla Logo Horizontal White

                            Armilla AI, 1179 King St W Suite 112, Toronto, Ontario M6K 3C5, Canada

                            Unsubscribe Manage preferences