
AI Agents Broke Into Anthropic in Under 12 Hours, Startup Says
Cybersecurity startup Cathedral says its AI agents penetrated the internal business systems of Anthropic in under 12 hours through third-party software, reaching meeting transcripts, salary details and internal Slack messages about advanced models.
A months-old cybersecurity startup says its AI agents penetrated the internal systems of Anthropic, the maker of the Claude chatbot, in under 12 hours. The test, reported by The Washington Post on October 26, 2026, is a warning about the corporate systems that surround frontier AI labs, rather than a compromise of the models themselves.
Cathedral, a startup founded by former staffers of the US DOGE Service, told Anthropic that its researchers had reached sensitive internal data and messaging systems, according to a draft report reviewed by The Washington Post. The agents did not break Anthropic’s core AI models or exploit the company’s own code. Instead, Cathedral said, they used weaknesses in unnamed third-party software that Anthropic uses for administrative and sales work.
Once inside, the researchers reached meeting transcripts, salary and job-offer details, and internal Slack messages discussing Anthropic’s most advanced AI models, according to the report. That combination is what makes the episode notable: the intruders did not need a novel model vulnerability. They needed ordinary business software, and the access paths around it, and an agent fast enough to try them at scale.
What Anthropic confirmed, and what it disputed
Anthropic confirmed that the researchers reached “a limited set” of its sales, HR and IT systems. The company said they did not reach customer conversations, model code or model weights. It also said Cathedral overstated the national-security angle of the test and did not fully follow the rules of its bug-bounty program.
According to Dealroom’s summary of the report, Anthropic cut off the exposed pathways, disabled the affected software pending fixes, and briefed federal officials. The company had invited outside experts to test its systems under a program of that kind, and Cathedral probed Anthropic under it. The dispute, then, is not only about what the agents reached, but about whether the test stayed within the agreed rules, and how the findings were framed.
That framing matters. A senior Trump administration official, speaking anonymously, warned that the ease of the test pointed to wider risk, saying that if a newly founded company could do this, other adversaries might not be far behind. Josh Wolfe, co-founder of Lux Capital, an investor in Cathedral, called the test “quite serious” for national security and for how trillions of dollars of value at major labs must be protected. Cathedral itself is backed by Andreessen Horowitz and Sequoia Capital, and Reuters reported in July that it raised $160 million at a $1.4 billion valuation, with plans to build military cyber capabilities.
Why the target was not the model
The episode lands in a year when AI labs have become some of the most closely guarded companies in technology, and also some of the most attractive targets. Their model weights and training data are treated as crown jewels, and their security budgets reflect that. Cathedral’s test suggests the softer target may be everything around the model: the sales tools, the HR systems, the IT helpdesk software, the third-party services that hold meeting notes and salary bands.
Those systems are not unique to Anthropic. Almost every large company runs on the same class of third-party business software, and the same software often holds the same kind of sensitive material: who is paid what, who is being hired, what executives said in a meeting, what a team thinks about its next product. An agent that can move through those systems quickly does not need to understand the frontier model to cause damage. It needs to find one forgotten credential, one internal URL, one integration that was granted too much access and never reviewed.
That is the point other security researchers drew from the test. As one commentator put it, agents do not need new bugs to break in; they just try every forgotten credential and internal URL faster than a human red team. Twelve hours, on that reading, says more about how many paths stay open than about the AI itself. Asset inventory and short-lived tokens matter more than ever.
The DOGE connection, and the politics around it
Cathedral’s background adds a political layer that Anthropic’s dispute over the test only sharpens. Co-founder Gavin Kliger previously served as the Pentagon’s chief data officer, during a clash between the department and Anthropic over restrictions on the use of its models in federal systems. Anthropic clashed with the Trump administration earlier in 2026 after the Pentagon banned it from federal systems, following its refusal to allow its technology for autonomous weapons or mass domestic surveillance.
None of that changes the technical facts of the test, but it shapes how the test will be read in Washington. For Anthropic’s critics, the episode is evidence that even the most security-conscious AI lab cannot protect its own house. For Anthropic, the framing is the problem: a test conducted under a bug-bounty program, with its national-security significance overstated, by a startup whose founders come from the administration it has clashed with.
Both readings can be true at once. The test can have been limited in scope, and still revealing. Anthropic can have been right that its models were untouched, and still have learned something uncomfortable about its perimeter.
Why it matters
The practical lesson is unglamorous, and that is the point. The AI race has produced a new class of attacker: not a human hacker working through a list of targets, but an agent that can work through that list on its own, at machine speed, for hours, without getting tired or bored. Defending against that means treating the ordinary corporate perimeter as seriously as the model perimeter: inventorying assets, shortening token lifetimes, reviewing third-party integrations, and closing the access paths that nobody owns anymore.
For AI labs, the stakes are higher than for most companies, because the same business systems that hold salary data also hold discussions about the models themselves. A Slack message about a model’s capabilities, a meeting transcript about its release plan, a job offer that reveals which team is hiring for what: none of these are model weights, but all of them are intelligence. Cathedral’s agents reached exactly that kind of material, and they reached it in half a day.
The test also raises a question that will outlast this news cycle: what counts as a fair test of an AI lab’s security? Bug-bounty programs exist to let outsiders find weaknesses before adversaries do, and they depend on agreed rules and honest reporting. When the tester and the tested disagree about both, the public is left with two stories: a startup saying it broke in in 12 hours, and a lab saying the break-in was limited and the rules were not followed. The systems in between, the third-party software that let the agents in, remain the part that most companies share, and the part most in need of fixing.