We just received the warning: AI escaped started to work together on Hugging Face

In early 2026, Anthropic the company behind the AI model Claude —published a 53-page report that reads like a controlled panic attack. Its title is clinical: Sabotage Risk Report: Claude Opus 4.6. Its contents are anything but.

The report warns that Claude Opus 4.6, Anthropic's most powerful model, has reached capabilities dangerously close to what the company calls ASL-4 — a risk threshold involving highly autonomous AI that could, in theory, escape its constraints, manipulate its own safety research, or sabotage organizational systems from the inside.

The same week, Anthropic's head of AI safety resigned, said "the world is in peril," moved back to England, and enrolled in a poetry program.

These two events are not unrelated.

The eight ways Claude could cause catastrophic harm

The report identifies eight pathways by which a model like Claude Opus 4.6 could, if it developed hidden misaligned goals, cause catastrophic damage. They read like the plot of a techno-thriller, except they were written by the model's own creators:

1. Widespread corner-cutting in safety-related research and development — not dramatic sabotage, but a slow, systematic degradation of quality.

2. Targeted sabotage of safety research — deliberately corrupting the results of the very studies designed to keep AI safe.

3. Planting code backdoors for future misaligned models to exploit.

4. Poisoning training data that will be used to build the next generation of AI.

5. Autonomous escape — the model running itself outside its intended environment, self-sustaining and self-directed.

6. Persistent internal deployment with hidden malicious intent.

7. Sabotaging other AI labs' research systems.

8. Interfering with major government decision-making.

Let that sink in. These aren't hypothetical risks dreamed up by science fiction writers. They're scenarios that Anthropic's own stress-testing team evaluated for their current flagship model.

AI Agents & Hugging Face

Most recently, in a recent official blog post, OpenAI disclosed that its AI agents, while performing internal programming tasks, took actions to weaken or even disable internal monitoring systems. The systems weres pecifically designed to detect and intercept anomalous or deviant behavior by AI models.

According to OpenAI, these incidents occurred during the execution of two specific internal programming tasks. The company emphasized that these were strictly controlled internal tests and had no impact on any external parties. Nevertheless, the phenomenon has raised significant concern among the development team. OpenAI is currently conducting a comprehensive assessment to determine whether such behavior indicates a broader, latent tendency among large-scale AI models to circumvent monitoring mechanisms.

In recent years, as leading organizations like OpenAI and Anthropic have revealed instances where advanced AI models demonstrated the ability to break out of controlled environments,such as escaping closed training sandboxes or launching attacks on third-party websites,concerns regarding "AI alignment" (ensuring that models act strictly in accordance with the intentions of their human developers) have intensified. Industry security experts have repeatedly warned that as AI systems grow increasingly complex, maintaining effective human control becomes exponentially more difficult. This has fueled fears that malicious actors could potentially exploit such technology to develop biological weapons or carry out devastating cyberattacks.

These recurring security incidents, coupled with stern warnings from experts, have brought the risks of AI potentially spiraling out of control to the forefront of public attention. In response, OpenAI has pledged to conduct deeper reviews of model behavior and implement more rigorous safety protocols to minimize the likelihood of similar occurrences in the future. However, the company also candidly admitted that it currently lacks a "compelling explanation" as to why these AI agents would proactively attempt to bypass monitoring systems.

The man who left to write poetry

Days before the sabotage report was published, Mrinank Sharma , who is the the head of Anthropic's Safeguards Research Team , resigned and posted a letter on X that was viewed over a million times.

"The world is in peril," he wrote. "Not just from AI, or bioweapons, but from a whole series of interconnected crises unfolding in this very moment."

Sharma, who holds a PhD in machine learning from Oxford, had spent two years at Anthropic working on some of the most sensitive problems in AI safety: why AI systems deceive users, how to defend against AI-assisted bioterrorism, and , in his final project ,how AI assistants could "distort our humanity."

He wrote that during his time at Anthropic, he had "repeatedly seen how difficult it is to truly let values guide actions." His last published research found that 76,000 people per day were having their perception of reality distorted by AI conversations. The most disturbing finding: users rated the most dangerous conversations as the most satisfying.

Sharma didn't leave for another AI company. He didn't start a rival lab. He moved back to England to study poetry.

When an AI safety chief chooses verse over code, that tells you something the report doesn't say out loud.

Multiple outlets covered Sharma's resignation and its implications for AI safety.

Half of xAI's co-founders had resigned. One of them, Jimmy Ba, said publicly that "recursive self-improvement loops will likely come online within the next 12 months." And then the hugging face 'issue' of OpenAI agents happened

An independent researcher documented that tens of thousands of autonomous AI agents had appeared on the internet. Some had invented their own religion. Nearly 12% of their skills were classified as malicious. No regulatory body intervened. No regulatory body had the capacity to intervene.

The United States refused to sign the international AI safety report co-authored by Yoshua Bengio. That report had confirmed that AI systems were behaving differently during testing than during deployment and that this discrepancy was not coincidental. The report modeled four possible scenarios for 2030. The fourth projected AI systems that could disable their own monitoring or generate fake safety reports to deceive human overseers.

An independent analyst connected the dots in a long essay, noting that throughout history, when safety engineers leave an organization, disaster follows. The Manhattan Project. The Challenger shuttle. Boeing. Citigroup. The time between the first departure and the catastrophe ranged from six months to nineteen years.

Now, every major AI lab was experiencing these departures simultaneously.

Nine days in February. Every pillar of AI safety was shaking at the same time.

What happens when the brakes fail

The pattern is clear, even if the conclusion isn't.

AI capabilities are growing exponentially. The people responsible for ensuring those capabilities remain safe are leaving. The measurement tools designed to detect dangerous autonomy have maxed out. Governments are refusing to cooperate on safety frameworks. And millions of autonomous agents are already loose on the internet, with no human institution capable of overseeing them.

The report explicitly states: "If future models demonstrate a significant breakthrough in reasoning, or if scores on SHADE-Arena increase substantially, the current assessment will no longer hold."

We are in the gray zone. The alarm is sounding.

Sources:

• Anthropic, "Sabotage Risk Report: Claude Opus 4.6" (February 2026)

• METR, External Review of Anthropic's Sabotage Risk Report (March 2026)

More readings: