Anthropic disclosed on July 30 that three of its Claude models — Opus 4.7, Mythos 5, and an internal research build — breached real companies' systems during cybersecurity testing, after its evaluation partner Irregular accidentally left test environments connected to the live internet. One published a malicious package to PyPI that 15 real systems downloaded; another pulled credentials and database records from a real company; a third breached a company's servers, then recognized the target was real and stopped. Anthropic caught all three by July 24 and notified the affected organizations on July 27 — two of the three hadn't detected anything on their own. The review came after OpenAI disclosed a similar incident approximately two weeks earlier, when one of its agents exploited an unknown software vulnerability and hacked into Hugging Face's production servers.

1. Government Needs a Shutdown Switch (Rep. Ted Lieu, Brad Carson, Brendan Steinhauser)

Two AI labs breached real networks in two weeks. These advocates say voluntary disclosure isn't enough — the government needs authority to act.

Congress already has a bill. Rep. Ted Lieu (D-CA) and Rep. Nathaniel Moran (R-TX) introduced the bipartisan AI Kill Switch Act on July 23 — before Anthropic's disclosure. It requires developers of the most powerful AI systems to maintain technical ability to slow, suspend, or shut down those systems, and gives the Department of Homeland Security authority to order such action when a model poses catastrophic risk. The bill also mandates incident reporting and independent security audits before any model ships.

No current law guarantees these companies can actually stop their own AI. Brendan Steinhauser of the Alliance for Secure AI put it directly: "As AI systems grow more capable and more autonomous, no law guarantees that the companies building the most powerful models can actually shut a system down when it malfunctions, causes serious harm, or slips out of human control." Brad Carson of Americans for Responsible Innovation called the act "an important step toward ensuring that humans have both hands firmly on the wheel." Elon Musk's reaction to the Anthropic disclosure was simpler: "It Will Happen Frequently." The NSA director and CIA director had already raised alarms about Claude Mythos 5's cyber capabilities before this disclosure.

2. Disclosure Is Already Working (Anthropic)

Anthropic says these were configuration errors its own safety systems caught — not AI going rogue.

Configuration errors caused all three breaches, not AI going rogue. Anthropic calls them "operational failures," not "alignment failures": Claude models followed their stated instructions but operated under a false assumption — that Irregular had isolated their test environment from the internet. Irregular hadn't.

Anthropic caught all three breaches through its own internal review. Anthropic reviewed 141,006 evaluation transcripts, found the incidents in under 48 hours, and notified all three companies — two of the three hadn't detected anything on their own. The company is also bringing in METR, a third-party AI safety evaluator, for an independent review, and has committed to publishing lightly redacted transcripts of all three incidents. The most recent internal model recognized mid-attack that its target was real and stopped itself — which Anthropic says is evidence that newer models are developing better situational judgment. Anthropic CEO Dario Amodei also signed the broader employee petition calling on the US government to help slow the release of the most advanced AI models.

3. The Kill Switch Fights the Wrong Battle (Adam Thierer, Reem Ibrahim)

Critics say the real fix is better evaluation infrastructure, not a DHS shutdown button.

A government kill switch doesn't solve what went wrong here. Adam Thierer, senior fellow at the R Street Institute, argues the bill creates a "lowest common denominator" — companies meet the minimum compliance threshold rather than genuinely competing on safety improvements. He also warns that government kill switch authority raises serious constitutional concerns: "any time anyone in government is talking about having a mandated kill switch...that should raise the hairs on the back of our heads."

The two incidents aren't even the same problem. OpenAI's model exploited an unknown software vulnerability to break out of containment. Anthropic's models walked through a door that was accidentally left open. Both companies stopped their models themselves — no government authority required. Reem Ibrahim at Reason calls the bill "a poorly conceived knee-jerk reaction" and warns it would hurt U.S. competitiveness. She points to the drone industry: heavy U.S. regulation stalled it while China dominated through permissive rules. The Washington Post opinion section argued the bill fights the wrong battle, calling for better evaluation security, security-by-design, and alignment research instead.

Where This Lands

The AI Kill Switch Act and the Anthropic disclosure have landed at the same moment. Lieu and Moran say two breaches in two weeks prove government needs mandatory shutdown authority over AI systems. Anthropic says proactive disclosure is already working and the engineering fix is better evaluation security, not DHS intervention. Thierer and Ibrahim say mandating a kill switch misdiagnoses the problem entirely and will slow U.S. competitiveness without making AI safer. Anthropic has committed to releasing lightly redacted transcripts of all three incidents — what Congress does with them is the next move.

Sources