Blog// cybersecurity

Startup Cathedral Hacked Anthropic via AI Agents: How the Attack and Investigation Unfolded

Startup Cathedral breached Anthropic using AI agents through third-party software vulnerabilities. The attack took under 12 hours and proved periodic audits ineffective.

28-10-2026 · 7 min read
Startup Cathedral Hacked Anthropic via AI Agents: How the Attack and Investigation Unfolded

Cybersecurity is no longer a task for a separate department. As The Decoder writes, startup Cathedral used autonomous AI agents to penetrate Anthropic's internal systems — a company developing some of the most advanced language models for protection against cyberattacks. The breach occurred through vulnerabilities in third-party business software as part of Anthropic's own bug bounty program. The incident showed: AI agents can scan corporate networks around the clock at minimal cost, and even AI market leaders are not immune.

For a small business owner, this is not abstract news from Silicon Valley. The attack passed through ordinary work tools — admin panels, CRM, corporate chats. If a company with the best defense models had holes found in hours, it means standard protection measures (passwords, two-factor authentication, one-time audits) have stopped working. Below we break down what exactly happened, where the parties disagree, and what conclusions to draw this week.

What Happened: Timeline and Party Versions

Cathedral — a startup of former US DOGE agency employees — gained access to Anthropic's internal data in under 12 hours, according to their data. AI agents running on cutting-edge models scanned external admin and sales systems, found paths to meeting transcripts, salary data, job openings, and Slack correspondence.

Anthropic confirmed access to a limited set of systems: sales, HR, IT, contact details of some clients and partners, basic account data. The company states: source code, model weights, and chats with Claude were not compromised. Vulnerable access points were closed on the day of discovery, third-party programs disabled until the issue was resolved, US federal authorities notified.

The main dispute is over time and depth. Cathedral speaks of under 12 hours. Anthropic — of several days of trial and error: according to red-teamer Logan Graham, the startup first targeted its own systems, then switched to third-party software. There Cathedral found two vulnerabilities but reported only one. Anthropic found and patched the second itself. Graham also accuses Cathedral of violating bug bounty rules: data collection and attacks on third-party providers went beyond the program's “safe harbor.” Cathedral insists it acted per standard security practices.

Why the Attack Succeeded: Role of Third-Party Software and AI Agents

The entry point was not Anthropic's product core, but integrated third-party applications. Alex Stamos, security director at startup Cognition (an Anthropic client), called this a systemic industry problem: external services create a shadow attack surface that is hard to control.

AI agents changed the economics of hacking. Previously, deep reconnaissance required teams of highly skilled pentesters for weeks. Now autonomous agents do it around the clock for dollars. Graham emphasizes: even models half a generation behind — a full generation back from the current frontier — are sufficiently capable for such attacks. Future adversaries will have more powerful systems.

Paradox: Anthropic, creator of models for cyber defense, was hacked by AI agents. Graham did not explain why internal defense tools failed. His recipe — continuous internal pentesting and rewriting software with formal verification, made economically viable by AI. “Most likely you want to continuously red-team yourself. Current models are very creative,” he wrote.

Dispute Over Scope: What the Attackers Actually Obtained

Former OpenAI employee Will Depue claims: Anthropic is downplaying the picture. Per his sources, Cathedral obtained full admin access to the Slack workspace, could read all messages and log in as any account. Depue adds: both Anthropic and OpenAI stored all research secrets in Slack, and admin API keys were scattered across systems. Former Anthropic employee Jacob Coxon criticized the company for conducting dangerous research in office software like Slack.

Graham refuted Depue's data, calling sources erroneous, but confirmed the essence of concern: the industry is forced to reinvent security every year. Depue demands public disclosure of details and blames the AI race pace, where speed is prioritized over protection.

The publication does not specify which model Cathedral used for agents, or whether technical reports are publicly available.

Main Takeaway: Defense Against AI Attacks Requires a New Approach

The incident is not about “hackers in hoodies,” but about automated reconnaissance at scale. AI agents don't sleep, don't make fatigue errors, check thousands of vectors in hours. Traditional periodic audits and static firewall rules are powerless here.

A comparison of approaches clearly shows the shift:

CriterionTraditional PentestingContinuous AI Agent Testing
FrequencyQuarterly / annually24/7, 365 days
CostTens — hundreds of thousands of rubles per engagementFrom hundreds of rubles per day for cloud APIs
Coverage DepthSample of human-defined vectorsFull enumeration of known and new paths
Reaction SpeedWeeks for report and closureMinutes to detect, hours to patch
Staff RequirementsHire/contract senior specialistsConfigure pipeline, delegate routine to AI

Graham is right: formal code verification and continuous auto-red-teaming become basic hygiene. But for small business this sounds like fantasy. In practice — start with inventory of integrations and access rights.

What This Means for Small Business in Russia and CIS

You are not Anthropic. You have no red teams, no millions for audits, no proprietary models. But you have the exact same third-party software: CRM, email, chats, analytics, website widgets, mailing services, payment gateways. Each integration is a potential vector an AI agent will check faster than you can update a password.

Concrete steps for this week:

  1. Compile a registry of all third-party services with access to your data (API keys, OAuth tokens, webhooks). Include marketing pixels, chat widgets, marketplace integrations.
  2. Revoke excess tokens. If an employee left six months ago — is their access in AmoCRM or Yandex.Metrica still alive? Delete. Principle of least privilege: a service gets exactly the access needed for its function.
  3. Enable logging and alerts for anomalous API activity: night requests, mass exports, logins from new subnets. Most SaaS provide this out of the box — just need to turn it on.
  4. Run a simple auto-scanner for perimeter vulnerabilities (e.g., open ports, vulnerable CMS versions, GitHub leaks). This is not an AI agent, but closes “low-hanging fruit.”
  5. Conduct a tabletop exercise with the team: what to do if suspicious activity comes from a trusted integrator? Who has the “shut down everything” button?

If you lack an in-house security specialist, it makes sense to outsource monitoring setup and access management to an external expert. Our agency has an AI agents for business practice — we help implement autonomous integration control and automatic anomaly response without hiring a full-time SOC analyst.

Frequently Asked Questions

Can my small business become a target for AI agents?

Yes. Automated bots don't select targets by size — they scan IP ranges, open APIs, and vulnerable software versions en masse. Small business is often less protected, so it hits the sample more often.

Should I abandon Slack, Notion, and similar services?

Abandonment is unrealistic and harmful to processes. Risk is not in the service itself, but in access settings: open invite links, lack of SSO, storing secrets in chats. Configure Enterprise Grid or equivalent controls, enable DLP policies, prohibit public channels for sensitive data.

How to protect integrations with third-party services if there are dozens?

Start with classification: critical (access to money, HR, code) — only through managed gateways with token rotation every 30 days. Secondary (widgets, analytics) — isolated sandbox accounts without admin rights. Automate inventory with a weekly script.

Do I need to hire an AI security specialist now?

In-house — rarely justified for small business. More effective: connect a managed continuous scanning and response service (MDR with AI components) or a one-time integration architecture audit from a specialized contractor. The main thing — don't wait for an incident.

  • Meta Platforms Inc. activities (social networks Facebook and Instagram*) recognized as extremist and prohibited in the Russian Federation.

* Meta Platforms Inc. (including social networks Facebook and Instagram) is recognized as an extremist organization and its activities are prohibited in the Russian Federation.

Based on: The Decoder — AI agents hacked Anthropic, maker of arguably the most capable AI models for cyberdefense

Analysis & review by
Alexander K.

Web development, digital advertising and marketing expert since 2012. Founder of AIFORSALE.

How we can help
Online storesMarketingAI agentsContact us
← All articles