AivexaNewsSearch
AI news for builders and product teamsChecked every hour

AI agents hacked Anthropic, maker of arguably the most capable AI models for cyberdefense

Collected Oct 28, 2026

Cybersecurity startup Cathedral used AI agents to break into Anthropic's internal systems, entering through external admin and sales software rather than Anthropic's own infrastructure, according to a draft report reviewed by The Washington Post. The work was carried out under Anthropic's own bug bounty program. Cathedral's researchers say agents running advanced models located paths to meeting transcripts, salary data, job postings and Slack messages in under twelve hours.

Anthropic confirmed that the researchers reached a limited set of sales, human resources and IT systems, including contact information for some customers and partners and basic user account data. The company said chats with Claude, source code and model weights were never exposed. It shut down the affected access points the same day it identified the breach, disabled the third-party programs until the issue was resolved, and notified U.S. federal authorities.

Cathedral was founded by former employees of DOGE, the U.S. government efficiency agency, and sells AI-powered security tools to government agencies and businesses.

The two organizations disagree on scale and duration. Cathedral puts the incident at under twelve hours; Anthropic says the attempted attack stretched over several days. According to Anthropic red teamer Logan Graham, the startup probed Anthropic's own systems for days before pivoting to third-party software, where it found two vulnerabilities but reported only one. Anthropic says it found and patched the second itself.

Anthropic also accuses Cathedral of incomplete disclosure in violation of the bounty program's rules. Graham said targeting third-party providers and pulling data went beyond the program's safe harbor provisions, while welcoming the discovery of the vulnerabilities. Cathedral maintains it followed standard security research practices.

The extent of the data access is contested. Former OpenAI employee Will Depue has claimed Anthropic is downplaying the incident, saying several sources he considers highly reliable told him Cathedral gained full admin access to the Slack workspace, could read every message and could log into any user account. Depue added that both Anthropic and OpenAI stored research secrets in Slack, with admin API keys scattered across their systems. Jacob Coxon, an Anthropic employee who left the company in September 2026, had previously criticized Anthropic for running dangerous AI research on office software like Slack. Graham said Depue's sources were wrong and declined to elaborate, but agreed with the core concern that the industry now has to reinvent its security every year. Depue called for disclosure of what Cathedral actually accessed and blamed the pace of the AI race, where executives prioritize speed over security.

Alex Stamos, chief security officer at Cognition, which uses Anthropic's models, told the Post he considered the findings significant. He noted that integrated third-party apps are a major problem across the industry and that Anthropic could have done a better job securing its external services.

For developers, the mechanics are the familiar shape of supply-chain risk. Business SaaS tools—admin consoles, CRM and sales platforms, messaging workspaces—are typically connected through OAuth grants and API tokens that grant broad scopes to whoever holds them. An agent that can enumerate exposed endpoints, read error messages, test credentials and chain small misconfigurations can work through that surface continuously. That is my reading of why the reported entry point matters more than the target: the weak link was connected software, not the model-serving stack.

Graham's warning is that attackers already run agents around the clock against external dependencies, and that models half a generation to a full generation behind the frontier are capable enough for this work, with stronger systems inevitably arriving. His recommended response is continuous in-house penetration testing and rebuilding software with formal verification, which he argues AI now makes economically viable. He advised red-teaming yourself constantly because current models are very creative.

A likely trade-off is that continuous automated red-teaming cuts both ways: the same agent capability that finds your exposures can be pointed at someone else's, at low marginal cost. Formal verification, meanwhile, is strongest on well-specified components and far harder to apply to sprawling SaaS integrations, so it addresses only part of the problem. Least-privilege scoping for third-party app grants, token rotation and logging of agent-driven access look like the more immediate controls, though the source does not prescribe them.

Anthropic is notable as the target because it markets itself on cyberdefense capability. That framing invites the obvious question of how far an AI-hardened lab should be expected to hold up when the intrusion arrives through a vendor, which the reported dispute over depth suggests is still unresolved.

Why it matters: Any organization exposing internal tools through third-party software is in scope, and security teams should assume adversaries can probe those connections continuously with agents. For product teams, the practical shift is treating every integrated app and API grant as part of the attack surface and auditing what data those connections can reach. Whether AI-driven defense meaningfully outpaces AI-driven offense remains an open question, and this episode is unlikely to settle it.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Cybersecurity startup Cathedral used AI agents to break into Anthropic's internal systems through third-party software as part of a bug bounty program. Anthropic confirmed access to sales, HR, and IT data but disputes how deep the breach went. The incident shows how vulnerable even leading AI labs are through their external software dependencies. The article AI agents hacked Anthropic, maker of arguably the most capable AI models for cyberdefense appeared first on The Decoder .