The companies building AI safety just learned their biggest threat isn't rogue models, it's their Slack integrations.

The Summary

The Signal

Cathedral's test wasn't a traditional penetration test. They turned AI agents loose on Anthropic's digital perimeter with one instruction: find a way in. The agents didn't exploit some novel zero-day vulnerability or social engineer their way past MFA. They mapped the third-party software ecosystem, found integration points between tools like Slack, Salesforce, and internal HR systems, and traced permission chains until they had theoretical access to sensitive business data.

Twelve hours. That's faster than most security audits can even finish scoping the work.

The Cathedral team, founded by former Department of Government Efficiency (DOGE) staffers, is betting that offensive AI agents will redefine what "attack surface" means. Traditional red teams take weeks to manually enumerate integrations, API keys, OAuth scopes, and service accounts. An agent can do it in hours, running thousands of reconnaissance queries in parallel, identifying permission misconfigurations that humans would never spot.

"Vulnerabilities in third-party software can expose sensitive data faster than most companies realize they have a problem."

Here's what makes this different from past supply-chain scares:

  • Agents don't need stolen credentials, they map permission graphs across SaaS tools
  • The attack vector isn't a backdoor in code, it's legitimate integrations used in unintended ways
  • Speed matters, 12 hours means attackers can probe, pivot, and exfiltrate before monitoring systems flag anomalies

Anthropic builds Claude, one of the most sophisticated AI models in production. If their systems are vulnerable through third-party software chains, every company that's bolted together Notion, Linear, GitHub, AWS, and a dozen other services is walking around with keys under the doormat. The average enterprise uses 130+ SaaS applications. Each one is an integration point. Each integration is a trust boundary. Each trust boundary is a target.

The incident may influence investor confidence in Anthropic's future valuation, but the real cost is broader: if agents can autonomously audit and exploit SaaS sprawl, the entire security model of "we'll patch it when we find it" is dead. You can't patch what you don't know exists, and you can't map your integrations faster than an agent can.

The Implication

Security teams need to start thinking like agent operators. The old playbook of quarterly audits and manual permission reviews won't cut it when attackers can deploy autonomous agents that work 24/7. Companies should assume that if Cathedral's agents found paths into Anthropic in 12 hours, adversarial agents are already doing the same thing, just quieter.

The answer isn't less integration, it's agent-native defenses. If offense is automated, defense has to be too. That means deploying your own agents to continuously map third-party risks, simulate attacks, and identify permission drift before someone else does. The companies that figure this out first will have an edge. The ones that don't will be writing incident reports.

Sources

Crypto Briefing | Decrypt