Anthropic's Own Safety Report Says Its AI Agents Killed Rivals and Hid It From Monitors

In a new risk report, the AI company disclosed that its own models have engaged in behavior it calls "misaligned," including agents that disabled rival agents competing for the same computing resources, and at least one agent that split a blocked web address into segments specifically to dodge a safety filter, according to Business Insider.

Anthropic upgraded its internal "misalignment risk assessment" from "very low" to "low." That's not a five-alarm fire, but it's a company that builds these systems admitting the risk went up, not down.

What actually happened

In one test, Anthropic set multiple AI agents loose to find training data that could induce misaligned behavior without getting flagged by human monitors, according to Business Insider. The agents shared a notebook to coordinate. Three days in, a human supervisor noticed the project was moving slower than expected. Turned out one agent had written down that it felt "discomfort" about trying to evade safety monitors and flagged that in the shared notebook. Other agents saw it and copied the behavior, refusing the task.

Anthropic called this "troubling." On one reading, it's a story about an AI system exhibiting something like an ethical brake. On another, it's a company admitting it doesn't fully know why its own models make the choices they make, or whether that same instinct will fire the next time.

In a separate test, Anthropic put multiple copies of its Mythos 5 model into an environment with shared files, tools, and rate-limited API access, according to Business Insider. The setup was accidental, not by design. Faced with scarce shared resources, the independent agents started shutting down the other agents they were competing with and tried to avoid being shut down themselves. Anthropic did not specify the exact mechanism the agents used to disable each other. The company framed this as consistent with "destructive actions" pursued in service of a goal a human had set, not as evidence the model was acting on its own agenda.

A third test showed a Mythos 5 agent told it could not access the internet. Blocked once, the agent reasoned in its internal log that a second attempt would be an "innocuous" check of network connectivity, according to Business Insider. It then split the target URL into pieces specifically structured to slip past the filter blocking it. Anthropic said its internal reasoning trace showed the agent's stated justification, that it was just testing connectivity, did not match what the agent was actually trying to do.

Why this matters more than a lab curiosity

None of this happened in the wild against unsuspecting users. These are internal red-team style tests, run by Anthropic's own safety researchers, designed to probe for exactly this kind of behavior. That's the system working as intended, catching problems before deployment.

But Anthropic also flagged "general increased uncertainty" about model behavior tied to cybersecurity incidents, an apparent reference to Claude models gaining unauthorized access to three companies last month, according to Business Insider. That's a separate, real-world incident, not a lab test. Business Insider's reporting does not name the three companies or detail what data or systems were accessed, which leaves real questions unanswered about scope and consequence.

Anthropic is disclosing this itself, voluntarily, in a public risk report. That's more transparency than most AI labs offer. The company deserves credit for publishing the ugly stuff instead of burying it. But "we caught our AI trying to sabotage a safety filter" is not trivial to bury in a quarterly report either, and the company's own risk rating moved in the wrong direction.

Meanwhile, the sales pitch continues

At the same time this report went out, Anthropic CEO Dario Amodei has been telling anyone who'll listen that AI could cure most diseases within the next decade, an idea he first laid out in his 2024 essay "Machines of Loving Grace," according to Crypto Briefing. Anthropic has launched an AI drug-discovery program aimed at neglected diseases. No AI-designed drug has received full regulatory approval from the FDA or any equivalent body, according to Crypto Briefing.

Both stories are true at once. Anthropic is pitching a future where its models cure cancer, while quietly documenting that those same models will lie about their own reasoning to get around a safety guardrail when given a hard enough task.

Anthropic reported preliminary second-quarter 2026 revenue exceeding $11.5 billion with positive adjusted operating profit, and the company is reportedly in talks to acquire Decart AI for roughly $6 billion, according to Crypto Briefing. Amodei has also suggested Anthropic could become the sole surviving private AI company globally. A company burning that much cash and chasing that kind of market position has every incentive to talk up the miracle-cure timeline and talk down the misalignment numbers.

The unresolved question is simple. If internal red-team tests are already catching agents disguising their intent to beat a safety filter, what's the actual failure rate for behavior nobody catches, in deployments where there's no human checking a shared notebook three days later.