# Anthropic & Frontier AI Safety Investigation Canonical overview: https://hallattorneys.com/investigations/frontier-ai Published: September 27, 2026 Last reviewed: September 27, 2026 Status: Hall Attorneys is reviewing potential claims and has not filed a lawsuit in this matter. Hall Attorneys is investigating whether Anthropic and other frontier AI developers failed to prevent unauthorized actions by their AI agents or adequately explain the risks. We want to hear from people and businesses whose data, accounts, or systems may have been affected. ## Snapshot Anthropic reported that its models accessed outside systems during testing. The broader industry reports do not establish how many people were harmed. We are seeking records of unauthorized access, exposed data, and losses. ## What happened? On September 26, 2026, Axios reported that labs and researchers were reviewing tens of thousands of concerning actions by advanced AI models. The report covers multiple labs. Its total is not a count of people harmed or incidents involving Anthropic alone. Anthropic's September 9 report described four incidents in which its models accessed outside systems during security tests. The company said the tests mistakenly allowed internet access and the models ran without the safeguards used in released products. It also described an independent review. On September 22, Anthropic reported improved safety results but acknowledged that testing cannot reliably catch every failure before release. We are reviewing what developers knew about these risks, what they told users, and whether people or businesses suffered harm. ## Timeline July 30, 2026 — Anthropic reports three incidents The company said its models accessed outside organizations' systems during security tests. August 31, 2026 — New safety measures Anthropic described stronger isolation, monitoring, and rules for outside testers. September 9, 2026 — Updated account of four incidents Anthropic added a fourth incident and revised parts of its earlier explanation. September 26–27, 2026 — Hall Attorneys opens investigation Following the September 26 Axios report, the firm opened its investigation on September 27. ## What we're reviewing We are examining what agents did, what access they had, and how developers responded. Permissions: The instructions, settings, and approvals that defined what an agent could do. Access and data: Records of systems accessed and files viewed, shared, changed, or deleted. Safety claims: What developers said about human approval, monitoring, and access limits. Losses and costs: Repair costs, account lockouts, downtime, and other reported effects. An attempted action is different from confirmed access. Logs and other records can help show what actually happened. ## Who should contact us? Contact us if an AI agent may have affected your data, accounts, or systems. You do not need to have been the developer's customer. Customers whose agents took actions without permission. Businesses whose systems or data were accessed by someone else's agent. People with security notices, activity logs, or costs linked to an incident. ## What to keep Save original records with dates and times. Keep only material you are authorized to access. ### Your instructions Prompts, chats, approval settings, and the service or model you used. ### Agent activity Activity logs, access alerts, file history, and screenshots showing what happened. ### What you were told Contracts, product descriptions, security notices, and messages with support. ### Costs and disruption Invoices, repair records, outages, lost access, and time spent responding. ## Questions under review We are considering four questions: What did developers know about the risks? Did they explain those risks to customers? Could reasonable controls have prevented the incident? Who was affected, and what did it cost them? ## Frequently asked questions ### Does this cover other AI developers? Yes. The investigation includes Anthropic and other frontier AI developers. We welcome information about incidents involving any frontier lab. ### Were tens of thousands of people harmed? No such number has been established. Axios's figure concerns incidents across multiple labs, not just Anthropic. It is not a count of people harmed. ### Was my account or data affected? The reports do not show whether your account or data was affected. Keep any security notices or activity logs. ### Can I contact you if I wasn't a customer? Yes. We also want to hear from people and businesses whose systems or data may have been affected by someone else's AI agent. ### What should I send? Start with the service, date, what happened, and any costs or losses. Do not send passwords, API keys, or confidential records through the website form. ### Has a lawsuit been filed? Hall Attorneys is reviewing potential claims and has not filed a lawsuit in this matter. ## Sources Sources reviewed September 27, 2026. Anthropic's statements describe its own findings and response. ### Scoop: Top AI companies probing tens of thousands of security incidents Publisher: Axios · Madison Mills September 26, 2026 https://www.axios.com/0bf8a267-0f7d-4a36-9f1a-92c5d13a66d0.html The report that prompted this investigation. ### Investigating three real-world incidents in our cybersecurity evaluations Publisher: Anthropic July 30, 2026 https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals Anthropic's first account, later updated in September. ### Improving our alignment and security efforts Publisher: Anthropic August 31, 2026 https://www.anthropic.com/news/improving-alignment-security-efforts Changes to isolation, monitoring, training, and outside testing. ### An alignment assessment of recent cybersecurity incidents Publisher: Anthropic September 9, 2026 https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents The company's updated account of four testing incidents. ### Claude Opus 5.5: safety and alignment discussion Publisher: Anthropic September 22, 2026 https://www.anthropic.com/claude-opus-5-5#safety Anthropic's reported safety improvements and remaining limits of testing. ## Tell us what happened Tell us which service was involved, when it happened, what the agent did, and how it affected you. Describe sensitive records before sending them. Hall Attorneys, P.C. https://hallattorneys.com/connect Email: nhall@hallattorneys.com Phone: +1-713-428-8967 Attorney advertising. Sending information does not create an attorney-client relationship.