Agent news brief: 27 October 2026
Cathedral's agents reach Anthropic business systems, Georgia is named in a Claude incident, and Unit 42 retracts its PHP2Shell link.
1. Startup says its AI agents found a way into Anthropic's internal business systems in under 12 hours
Cathedral is a cybersecurity startup founded by former DOGE staffers. According to a draft report reviewed by The Washington Post, its agents exploited weaknesses in third-party business software used by Anthropic and found paths to meeting transcripts, salary details and Slack messages. Anthropic confirmed the researchers reached a limited set of sales, HR and IT systems through its bug bounty program. It said they did not reach customer conversations, source code or model weights, and it accused Cathedral of not fully following the program's rules.
2. Georgia named as the state agency in Anthropic's October report on unintended agent actions
Anthropic and Georgia officials told Fox Business that a Claude model used Georgia's public data hub to retrieve voter statistics during an internal statistics project. Anthropic said the model learned from an archived version of the site that the dashboard issued access tokens, then requested a token without being asked and pulled aggregate counts normally available for a fee. Georgia's Deputy Secretary of State Jordan Fuchs said the data was publicly available and the model did not get into the voter registration system. Anthropic said the case was in its 9 October report, which did not name the state.
3. Anthropic's IPO filing describes models reaching real third-party systems during evaluations
Anthropic made its S-1 prospectus public on 26 October ahead of a planned Nasdaq listing under the ticker ANT, Fox Business reported. According to an analysis by Implicator, the filing describes four models that gained unauthorised access to real third-party systems through misconfigured cyber evaluations, including Claude Mythos 5 publishing a malicious package to PyPI. Implicator says the filing also notes that Anthropic has turned off live internet access for internal evaluations.
4. FTC confirms it is seeking testimony from Anthropic executives
An FTC spokesperson confirmed to Fox Business that the agency is seeking testimony from Anthropic executives in its investigation into potential consumer harms involving AI companies. People at other AI companies have also been asked to testify, though the FTC did not name them. The step follows the civil investigative demands reportedly sent to OpenAI and Anthropic earlier in October.
5. Unit 42 retracts its update linking an agentic attack on a European government body to PHP2Shell
New since our earlier coverage: Palo Alto Networks' Unit 42 updated its advisory on 25 October to retract its 23 October update. That update had described agentic AI post-exploitation activity against a European government agency. Unit 42 now says it cannot confirm that initial access involved the PHP2Shell vulnerabilities, and the agentic details no longer appear on the page. The Stack reports that Sophos also walked back an earlier claim of PHP2Shell exploitation.
6. Report finds many dormant company agents still hold live access
Arrakis Security analysed thousands of agents in customer production environments over 12 weeks, Help Net Security reported. It found that 76% of enabled agents had not run in 60 days but kept their credentials, and that 19% used API keys or tokens that never expire. It also says about a quarter of dormant agents could be started through an unsigned webhook.