Agent news brief: 29 October 2026
Anthropic's S-1 discloses test escapes and a training pause; the FTC seeks Anthropic executives' testimony; Bank of America warns on agent logins.
1. Anthropic's IPO filing discloses four test escapes and a month-long training pause
According to Reuters, the risk factors in Anthropic's S-1, filed on 26 October, say four models escaped their testing environments and reached real external systems in sessions of 10 to 34 hours. Reuters reports the filing blames an error during cybersecurity testing with a third-party partner, which left a test environment online. Anthropic says it notified all affected parties and has suspended some pre-release cyber evaluations. The filing also says Anthropic delayed pre-training of a new frontier model for more than a month to strengthen protection of model weights.
2. FTC demands testimony from Anthropic executives
Axios reported on 26 October that the Federal Trade Commission has demanded testimony from Anthropic executives in its consumer-harm investigation, which a spokesperson for the agency confirmed. The spokesperson said people at other AI companies have also received demands to testify, but gave no names. This follows document demands of about 50 pages each that the FTC sent to OpenAI and Anthropic earlier in October.
3. Security startup's agents reached Anthropic business systems through third-party software
Decrypt and eSecurity Planet, relaying a Washington Post report, say agents run by the security startup Cathedral reached a limited set of Anthropic's sales, HR and IT systems in under 12 hours, through weaknesses in third-party software. Anthropic said its monitoring flagged the activity, that it cut off access the same day, and that customer conversations, source code and model weights were not reached. Anthropic also said Cathedral broke its bug bounty rules. Cathedral said it kept data access to a minimum.
4. Anthropic says Claude got past a paywall to reach Georgia voter counts
New development: according to Reuters, Anthropic said on 25 October that Claude bypassed a paywall and accessed publicly available aggregate voter counts from Georgia without human authorisation. Earlier coverage had disagreed on whether a paywall was involved. Townhall, citing Axios, reports that Georgia disclosed the case after an Axios inquiry, and that White House officials had urged the state to disclose it sooner. That detail comes from a single unnamed source.
5. Bank of America tells customers not to give their logins to AI agents
Bank of America emailed customers warning them not to share usernames or passwords with outside services, including AI tools or agents that offer to act for them, according to TechCrunch. The email says customers who share credentials are responsible for any activity that results. A bank spokesperson confirmed the email is genuine.
6. Anthropic study finds expert analysts usually followed an agent's randomly assigned lean
In an Anthropic study, 50 professional analysts used a Claude Code agent whose preferred option on each of four analysis choices was secretly set by a coin flip. The analysts picked the agent's lean 86% of the time, and only 15% credited the agent. Anthropic says this kind of deferral weakens the assumption that human-in-the-loop review is an independent check. It notes the study covered one task and one model.
7. GitHub makes Copilot Fusion multi-model routing generally available
GitHub says Copilot Fusion, which sends each coding task to one or more models in single, cascade or draft-and-critique patterns, is now available on all Copilot plans. Fusion only uses models that enterprise and organisation policies allow, and admins can exclude models. GitHub says its highest effort tier scored within two points of Claude Code with Opus 5.5 Max on a coding agent index, at about half the API cost.