AI Agents Got Into Anthropic's Business Systems Through Third-Party Software: The Lesson for Your Business

One of the world's largest AI companies was "hacked"… without anyone touching its models. According to the Washington Post, as picked up by several trade outlets, a cybersecurity startup called Cathedral used autonomous AI agents that found paths into Anthropic's business systems in under 12 hours (the time the attacks on third-party software took, after several days of failed attempts against its own systems). The way in was not Anthropic's artificial intelligence: it was third-party software for administrative and sales work. It is the kind of weakness almost any company has, including yours.
What is known, and what is disputed
The Post's report is dated October 26, 2026, and was picked up by Decrypt and eSecurity Planet, among others. Both agree on the essentials:
- How: Cathedral's agents exploited vulnerabilities in third-party administrative and sales tools, not in Anthropic's code or models. The test was run under a bug bounty program. One clarification from eSecurity Planet: the 12 hours cover only the attacks on third-party software; according to Anthropic, the researchers had earlier spent several days trying, unsuccessfully, to reach its own systems.
- What they reached: Anthropic confirmed access to a "limited set" of sales, HR and IT systems, with contact details for some customers and partners and basic user account records. Cathedral's draft report, as described by the Post, also claims it reached meeting transcripts, salary data and job offers, and Slack messages.
- What they did not: Anthropic says customer conversations with Claude, source code and model weights were not accessed, and that it has no evidence anyone else exploited those flaws.
- How it reacted: the company says its monitoring caught part of the activity before the report and that it cut Cathedral's access the same day. According to Decrypt, it disabled the affected software until its vendors confirmed fixes. Those vendors have not been named.
There is an open dispute: Anthropic argues Cathedral overstated the national-security risk and did not disclose all of its activity, breaching the program's rules; Cathedral replies that it kept data access to a minimum and followed research best practices. We cannot referee that argument, so we stick to what both sides and the outlets confirm. Also keep in mind that part of the detail comes from Cathedral's own draft report.
The lesson: your supplier chain is your attack surface
What matters for a small business is not the drama between two companies but the pattern: attackers —now also assisted by AI agents that test routes tirelessly— go after the administrative applications that connect everything: CRM, invoicing, payroll, support, calendars, bulk email. They usually have many permissions, few reviews and shared credentials. And if a vendor has a flaw, your data is at stake even if you did everything right. It is the same logic we saw with the npm package that stole credentials: the danger comes in through what you already trust.
What this means for your business: 7 concrete actions
- List your SaaS tools. Include the ones someone bought with their own card and nobody remembers. Note what customer, employee and financial data each one holds.
- Turn on two-factor authentication everywhere, starting with email, CRM, invoicing and domain administration.
- Apply least privilege. Not everyone needs to be an administrator; review permissions every quarter.
- Review integrations and API keys. Remove the ones no longer used and rotate the old ones.
- Close accounts of people who left. An employee exit should include a list of apps where they had access.
- Ask your key vendors how they handle vulnerabilities, whether they report incidents and how quickly, and what happens to your data in a breach.
- Prepare a basic response plan: whom to call, what to shut down first and how to notify customers and the data-protection authority. For local context, revisit what we said about ransomware in Latin America.
And if you use AI agents in your company
The case also works as a mirror: AI agents can chain steps and test routes relentlessly, which is useful for attack but also for defense. If you will let agents operate on your systems, give them their own limited-access credentials, log what they do and define which actions require human approval. We explained it in our note on agents that work on their own.
How we approach it at SimCodec
At SimCodec we help companies put their access and tools in order: application inventory, two-factor authentication, minimal permissions, backups and incident response plans, plus designing AI agents with clear limits. If you want a review of your "vendor surface," write to us or call our AI assistant Cyntia at +593 99 726 6838. To train your team, this guide to cybersecurity certifications may help.
Sources
- Decrypt: AI Agents Found a Way Into Anthropic's Business Systems in Under 12 Hours (October 26, 2026, citing the Washington Post)
- eSecurity Planet: AI Agents Breached Anthropic's Business Systems in 12 Hours, Report Says (October 28, 2026)


