Ex-DOGE researchers breach Anthropic systems in under 12 hours using AI agents

Ex-DOGE cyber researchers penetrated Anthropic's systems in under 12 hours using AI agents, gaining access to Slack channels where sensitive model discussions take place. The breach was a white-hat security exercise conducted under safe harbor, not a malicious attack, and came weeks after Anthropic whistleblower Joseph Coxen leveled separate criticism at the company.

The timing was deliberate. The researchers apparently planned to drop the story on Anthropic's IPO day for maximum impact, though the segment doesn't clarify whether that timing materialized or was preempted by other news.

The incident underscores a structural gap in how AI labs incentivize security testing. Bug bounties from companies like Anthropic and OpenAI typically max out in the six to seven-figure range, but the economic stakes of a critical infrastructure compromise are vastly higher. A breach of banking systems or the electrical grid could trigger GDP-scale damage comparable to pandemic-level disruption — potentially hundreds of billions in losses. Against that arithmetic, current bounty budgets look undersized.

One proposal floated is a $20 billion government-backed bug bounty program modeled after Intel's past deals, with eligibility restricted to major companies and security firms. The mechanism: companies like Meta could allocate spare GPU compute capacity to participate, hunting vulnerabilities not just in their own systems but potentially in the small businesses that advertise on their platforms, converting security work into government payouts at scale.

The counterargument is chaos. Opening a massive bounty to freelance researchers creates moral hazard. Demand for tutorials has already emerged — people in Miami are reportedly selling courses on white-hat bounty techniques using open-source models. If incentives climb high enough, researchers with no experience handling sensitive data could be drawn in, and the distinction between white-hat disclosure and straight theft of valuable information (like medical records) becomes contestable.

The structural problem remains: either you keep bounties small and rely on individual companies' PR incentives to drive security testing, or you scale incentives and accept spillover risks you can't fully contain.