On Saturday 7 November, Anthropic began emailing its business customers with unwelcome news: limited business contact and account information may have been subject to unauthorised access by a third-party cybersecurity research company. The email does not name the company. But the description matches the one startup Anthropic confirmed in October had reached its internal systems: Cathedral, which says its AI agents found a way in within 12 hours.
Five days before the email went out, OpenAI announced a partnership with Cathedral for its next internal security sprint. It is the startup CPO Magazine reports Anthropic is refusing to pay a bug bounty.
What the email actually says
The notice, titled “Notice Regarding Your Organization’s Account Information”, was sent from Anthropic’s Claude no-reply address. Security researcher Jamieson O’Reilly shared three screenshots on X at 10:50 UTC on 7 November, CellCog reported, and the post drew more than 1,000 likes within a day, MLQ.ai reported.
For organisations with enterprise Claude accounts, the email lists account ID, organisation ID, plan tier, usage and spend metrics among the fields that may have been viewed. It says the information “was limited to business contact and account information”. Conversation content with Claude, passwords, and payment card and bank account information were not involved, according to the notice.
Anthropic says it disabled the third party’s access once it became aware of it, worked to fix the vulnerabilities it identified and opened an investigation. “The cybersecurity firm has certified in writing that it has deleted the accessed data and will not use it,” the notice says. Anthropic adds that it found no evidence the data was altered, deleted or misused, and that it reported the incident “to the relevant Korean authorities”. The notice does not say why Korean regulators in particular were notified.
Two things the email does not do: it does not name the firm, and it does not say how many organisations or people were affected. Anthropic says it “cannot confirm whether any particular person’s information was viewed”, and that the notice is sent “out of an abundance of caution should you wish to notify individuals in your organization”. In other words, it is leaving it to each customer organisation to decide whether to warn its own staff.
O’Reilly, in a follow-up, asked who the firm was, whether Anthropic had reported the access to law enforcement, and how the company could say it cannot confirm whose data was viewed while also saying recipients face no added risk. The notice answers the second question only in part: it says Anthropic reported the incident to the relevant Korean authorities, without naming them or saying whether law enforcement was notified. It also tells recipients there is “no reason to believe you are at heightened risk”, a passage O’Reilly quotes though he did not screenshot it, which sits uneasily beside the admission that Anthropic cannot confirm whose information was viewed. The notice leaves the firm unnamed. Anthropic has not publicly said whether the notices concern Cathedral, and no public Anthropic post about the notification has been found. The firm is described in the email as “conducting a security testing exercise on its own initiative and without our authorisation”. The match with Cathedral rests on that description alone.
The October test that fits the description
The Washington Post reported on 26 October that Cathedral, a months-old cybersecurity startup founded by former members of the US Department of Government Efficiency, said its AI agents found routes into Anthropic’s internal business systems in under 12 hours. The entry route was weaknesses in third-party business software Anthropic uses for administrative and sales work. Claude itself was not the route. Cathedral says it tested under Anthropic’s programme inviting outside researchers to probe its systems, and disclosed its findings to Anthropic earlier in October. The vendors’ names and the exact flaws were withheld.
Anthropic confirmed to the Post that Cathedral reached “a limited set” of its sales, human resources and IT systems, holding business contact information for some customers and partners plus some basic user account records. It said customer prompts, Claude responses, source code and model weights were not accessed. Its monitoring flagged some of Cathedral’s activity before the startup reported it, and Anthropic shut off access the same day, disabled the affected business software until the vendors confirmed fixes were ready, and briefed federal officials. It said it has “no evidence that anyone exploited these vulnerabilities outside of Cathedral.”
The two sides disagree on how far it went. Cathedral’s draft report describes paths to meeting transcripts, salary and job-offer details, and Slack messages discussing Anthropic’s advanced AI models. Anthropic’s own statement does not confirm those three categories. It confirms only that customer prompts, Claude responses, source code and model weights were not accessed, and says nothing specific about those three. It also says Cathedral overstated the “implied national security implications” of its findings and did not fully follow the bug bounty rules, because it failed to disclose all of its activity. Cathedral’s position, as quoted in its draft report, is that “data accesses were kept to the absolute minimum in accordance with security research best practices.”
Alex Stamos, chief information security officer at AI coding startup Cognition and formerly director of the Stanford Internet Observatory, reviewed Cathedral’s findings at the Post’s request. “They found something big here,” he said, and the episode showed “a gap” in how Anthropic manages its use of outside software providers. “This is a really hard problem, but Anthropic could have done a lot better.”
Cathedral was co-founded by Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein, all former DOGE staffers; Kliger served as the Pentagon’s chief data officer. The startup raised $160 million in July at a $1.4 billion valuation, with Andreessen Horowitz and Sequoia Capital leading. A senior Trump administration official, speaking anonymously, told the Post the report had circulated within the government and raised national security concerns: “If a bunch of kids who have just started a company can do this, then potentially, the Chinese aren’t far behind.” The same official said three federal agencies had recently contracted with Cathedral. That framing is the official’s claim, disputed by Anthropic, not the company’s position.
Three accounts, one gap
Put the three accounts side by side and the gap is the story. Cathedral claims more than Anthropic confirms, and the customer email says more than either of them said in October.
| Cathedral claims (via the Post) | Anthropic confirms (statement to the Post) | The 7 November email adds | |
|---|---|---|---|
| Time | Routes found in under 12 hours | Access shut off the same day; monitoring flagged some activity first | Customers told 12 days after the Post report |
| Systems | Internal systems via third-party admin and sales software | “A limited set” of sales, HR and IT systems | No additional systems identified; enterprise account information fields listed |
| Data | Meeting transcripts; salary and job-offer details; Slack messages about advanced AI models | Business contact information for some customers and partners; basic user account records | Account ID, organisation ID, plan tier, usage and spend metrics |
| Explicitly not reached | Customer prompts; Claude responses; source code; model weights | Conversation content; passwords; payment card and bank account information | |
| Outside exploitation | No evidence anyone outside Cathedral exploited the weaknesses | Firm certified in writing it deleted the data | |
| Count | None given | None given | None given |
Our reading: the email is the most specific account yet of what may have been exposed, and it is the only one that names the kind of data enterprise customers should worry about. It names no firm and no count.
One number in the story needs a correction. The “under 12 hours” figure covers only the vendor phase. CPO Magazine reported that Cathedral’s researchers first spent several days attempting to breach Claude itself, apparently to no avail, before turning to the vendors, where they found success in under half a day. Days of failed direct attempts came first. An attacker after theft, not a bounty, would likely have skipped straight to the vendors.
No bounty from Anthropic. A partnership from OpenAI.
Anthropic is refusing to pay Cathedral any bug bounty, CPO Magazine reported, because the researchers did not disclose all of their activity and failed to follow the programme rules. CPO Magazine reports the programme pays $100 to $550,000 depending on severity. Cathedral has responded by saying it intentionally did not go as far as it could have.
On 2 November, five days before Anthropic’s email, OpenAI announced a partnership with Cathedral for its next internal security sprint, whose researchers “will use OpenAI’s models alongside our teams to challenge our defenses”. Its last sprint dedicated 25% of its product engineers to finding and fixing weaknesses. OpenAI also raised its bug bounty rewards, more than quadrupling the midpoint of its reward ranges for medium, high and critical findings on average: a critical vulnerability in ChatGPT, Codex or the API Platform now earns up to $30,000, and exceptional critical findings remain eligible for up to $100,000, rising to $172,000 with bonuses. Its own example: a zero-click account takeover in ChatGPT that once paid $6,500 now pays $51,577. Neither OpenAI post mentions Anthropic, and no fee, start date or scope for the sprint has been given.
The wider point is not about one startup. CPO Magazine notes Verizon’s DBIR found 48% of data breaches involved a third-party services provider, up 60% on the prior year, and Kroll’s Data Breach Outlook 2026 found 69% of affected individuals in the incidents it handled were impacted by a third-party breach. Jacob Krell, senior director of secure AI solutions and cybersecurity at Suzu Labs, put it bluntly: “Anthropic is about as well-defended as a company gets, and its crown jewels never moved, no model weights, no source code, no customer conversations with Claude. What got hit was the layer it bought instead of built. Security maturity is capped by the weakest vendor you integrate with, and no amount of in-house engineering changes that.”
What to watch next
Whether Anthropic confirms the 7 November email concerns Cathedral. Whether Cathedral publishes its full report, and whether it names the vendors or the flaws. Whether any vendor named in a future disclosure issues its own statement or patch notice. And whether OpenAI’s sprint, with Cathedral inside it, finds anything. Its security team will be at PROMPT2PWN in Oakland on 20 and 21 November, where researchers can test targets in person.
If any of those appear, this article will be updated.


