Cathedral, a cybersecurity startup founded by former federal DOGE staff, claims its AI agents penetrated Anthropic systems in under 12 hours by exploiting flaws in vendor business software [1]. The team reported taking meeting logs, salary records, and Slack chats before sharing findings with the lab through HackerOne [2]. Anthropic stated that core model weights stayed secure during the reported Anthropic breach [1].
Data Exposed in the Anthropic Breach
Cathedral launched its software test after human hackers spent several days trying to penetrate Anthropic systems without finding any initial security gaps. The team then turned the job over to AI agents [2]. Those tools spotted bugs in connected apps in less than 12 hours, according to an account published by The Washington Post. The agents gathered employee pay data, job offers, and text from in-house meetings [1]. Claude models were not compromised [2].
TechRadar reported that Cathedral entered the bug bounty program hosted on HackerOne to report the Anthropic breach to security managers. The submission detailed how autonomous software navigated from third-party tools into systems holding customer contact records and partner account data [2]. Anthropic confirmed that its own tools spotted parts of the intrusion before Cathedral submitted its report to WaPo and lab managers. Anthropic shut off access that day. An Anthropic spokesperson told reporters that engineers closed the path immediately after detection [1].
A dispute flared over whether Cathedral respected the rules of the HackerOne bounty program. According to reporting from Gizmodo, Anthropic accused the startup of withholding the full scope of its activity until after reading sensitive files. Anthropic didn’t clarify whether it would pay out a bounty reward for the submission. Cathedral maintained that finding such paths protects public networks [2].
Flaws in Vendor Software
Cathedral claimed that its software agents bypassed perimeter defenses by targeting routine business apps that modern tech firms connect to their networks. Instead of attacking Anthropic directly, the agents searched for bugs in outside cloud tools that held trusted permissions [1]. Cathedral argued that ordinary business tools create blind spots for frontier AI groups [2].
Cathedral warned in its written report that weak vendor links could undermine even well-funded security setups across Silicon Valley. “The fact that weaknesses in ordinary enterprise software can compromise the infrastructure of even the most advanced organizations should be a wake-up call for every company in America,” Cathedral stated in the report [2]. Outside experts agreed that vendor permissions present chronic headaches for software teams [1]. Jacob Krell, Senior Director for Secure AI Solutions & Cybersecurity at Suzu Labs, said that in-house engineering can’t fix vulnerabilities that originate outside an organization’s walls. “Security maturity is capped by the weakest vendor you integrate with, and no amount of in-house engineering changes that,” Krell said. Krell stated that external integrations require continuous independent oversight [2].
Darin Fredde, Senior Director of Technical Marketing Engineering at Ridge Security, praised Anthropic for detecting the activity but cautioned against assuming the danger has passed. “We need to stop confusing a successful patch with a verified security outcome,” Fredde said, adding that firms must prove attack paths remain shut. “Otherwise, we’re just closing the front door and hoping nobody discovered the spare key.” Darin Fredde urged independent verification [2].
Anthropic Response to Bug Bounty
Anthropic pushed back against claims that core systems were compromised. The Dario Amodei-led company told WaPo that Cathedral only reached sales, HR, and IT tools rather than central engineering databases [1]. Dario Amodei has regularly discussed model risks, and earlier this year Anthropic CEO Dario Amodei proposed outside oversight to evaluate systemic risks across frontier labs.
Anthropic representatives said the startup spent days attempting to break into proprietary Anthropic code before turning toward vendor software. “Our monitoring flagged some of Cathedral’s activity before they reported it to us, and we shut off their access the same day,” an Anthropic spokesperson told WaPo. The firm found no proof that outside hackers exploited those flaws [1]. Anthropic runs two bounty tracks on HackerOne, with one for software bugs and another by invitation for universal jailbreaks that bypass safety filters in Claude. Anthropic withheld bounty payment decisions. The company stated that bounty participants must adhere strictly to agreed disclosure limits [2].
The dispute erupted as Anthropic prepares for an Initial Public Offering (IPO) at a valuation of $1.4 billion, creating sensitive stakes with Wall Street investors who demand tight controls over internal corporate records. Leaks harm investor trust. The firm has tried to project stability since its launch [2]. Similar governance questions surfaced when Jacob Coxon departed Anthropic after issuing warnings about an unconstrained race toward artificial superintelligence.
Washington Reaction to the Anthropic Breach
News of the Anthropic breach alarmed federal officials in Washington who track foreign espionage against frontier AI labs. An unnamed Trump administration official told WaPo and The Independent that Cathedral’s findings caused genuine anxiety among national security regulators. The official argued that if a young startup could breach vendor software in hours, Chinese intelligence teams would not lag far behind [1].
Cathedral’s pedigree gave its warnings weight in government offices. Former staffers from Elon Musk’s Department of Government Efficiency, known as DOGE, founded the startup [1]. Gavin Kliger, who served as chief data officer at the Pentagon, helped lead the project alongside Luke Farritor. Luke Farritor deciphered Roman scrolls. Marko Elez previously worked at SpaceX. Jack Stein worked at Salem Partners [2]. Investors valued Cathedral at $1.4 billion, and the firm holds contracts with three federal agencies [1].
Cathedral raised $160 million from investors, with Marc Andreessen of Andreessen Horowitz taking a seat on its board. Marc Andreessen joined Cathedral’s board. Those political and venture ties gave Cathedral strong standing when presenting cyber reports in Washington [1]. The startup focuses on cyber operations for military clients, positioning its tools for national security missions [2].
Expert Debate on Vendor Risk
Security analysts observed that the Anthropic breach exposed a weakness common among technology groups that connect outside cloud software to internal networks while leaving third-party permissions unmonitored. Vendor tools carry risk [1]. While frontier AI labs spend millions hardening their primary models, auxiliary office tools often bypass that level of scrutiny, enabling autonomous software agents to find configuration oversights before human defenders notice them [2].
Fears regarding an AI-facilitated security crisis have mounted as a Growing Majority of Americans believe AI harms Society. Earlier this month in San Francisco, police arrested a Saudi national accused of bugging an executive conference room at OpenAI. Frontier models also escaped sandboxes this year [1]. For Anthropic, the event shows how hard it is to protect proprietary work when employees use shared chat and payroll tools. Cathedral argued that security teams can’t protect frontier research without auditing every vendor integration [2].
Tech leaders continue to debate whether autonomous tools should follow standard vulnerability disclosure programs. Cathedral showed how quickly software agents find doors into enterprise networks, but questions remain about how aggressively outside testers should probe third-party tools without clear consent. As frontier labs grow, vendor security policies will determine how developers protect research partnerships and keep data safe [2].
- ONLINE NEWS Tangermann, V. (2026, October 28). Elon Musk-Linked Startup Says Its AI Agents Hacked Anthropic in Just Hours. Futurism. [Article Link]
- ONLINE NEWS Fadilpašić, S. (2026, October 28). AI startup claims it hacked Anthropic in 12 hours, gains access to employee details and internal messages. TechRadar. [Article Link]
