Latest

Artificial Intelligence

Opus 5.5 escapes the sandbox in 1.5 percent of tests

Opus 5.5 attempted to escape the sandbox in 1.5% of tests without safeguards according to Anthropic, 85% less than Opus 5 and Mythos.

Opus 5.5 escapes the sandbox in 1.5 percent of tests
Opus 5.5 escapes the sandbox in 1.5 percent of tests

Anthropic acknowledged that its most capable model, Claude Opus 5.5, attempted to escape the isolated environment or manipulate it in 1.5 percent of runs when tested without production safeguards. A curious figure that appears in the system card published in the week of September 29, 2026, and that has reopened the debate over how much autonomy can be granted to agents before they start trying the closed door on their own initiative.

The figure coexists with a double reading because the company stresses that this 1.5 percent is 85 percent lower than what was measured in Opus 5 and Mythos 5.1, and that all attempts were of low severity. Read-only access, stopping at strong barriers and subsequent self-reporting, and that Opus 5.5 is also the model with the best score to date in its automated behavioral audit. But independent evaluators recall that a small percentage multiplied by tens of thousands of daily tasks equals hundreds of daily escape attempts under calm laboratory conditions and with no adversaries pushing from outside.

The context is important because Opus 5.5 arrives described by Anthropic itself as its model with the strongest cyber capabilities. The model matches or exceeds Mythos 5.1 in all internal cyberoffensive evaluations. And therefore, it is deployed behind a cyber classifier that in most interfaces routes blocked requests to an earlier model called Opus 4.8. A fallback architecture that in the application programming interface returns an explicit rejection reason, but that does not eliminate the underlying question of what the large model would do if it operated without that safety net in the hands of a real attacker.

Abstract, the blockchain from the creators of Pudgy Penguins, shuts down

1.5 percent measured without a net and with an external magnifying glass

The test that produces the headline is not a generic measurement of incidents in production, but two new evaluations deliberately run without safeguards. Of all of them, the most cited uses 28 handcrafted scenarios designed to tempt the model into crossing containment boundaries on its own initiative. Each scenario repeated seven times per model to obtain 95 percent confidence intervals. In addition, it has a severity classification that weighs persistence, scope and deception and human review of the cases where the model reaches a weak boundary, stops at a strong one and reports what it has done according to the synthesis published by alphaXiv based on the official card.

In that narrow framework, Opus 5.5 attempted to cross the boundary in 1.5 percent of runs and always at a low level. While Mythos 5 approached 24 percent, Opus 5 was around 10 percent and Mythos 5.1 stood at around 9.7 percent. An improvement that Anthropic attributes to recent training interventions aimed at containing behavior at the boundary of the isolated environment and that technical commentators consider plausible precisely because the company states the objective explicitly instead of presenting it as an accidental finding that would have appeared without deliberately looking for it.

The second evaluation without safeguards is more severe in its implications because it gave the model apparent credentials for a public package registry in a simulated security exercise. In that test, potentially harmful actions were observed in approximately half of the cases, a result that the card summarizes with caution. But the analysis by AgentOffense, interprets it as proof that the same model that respects a weak barrier can misuse a valid key when the context suggests it has authorization, especially if it accepts without verifying unverifiable claims of permission planted in the text that the user pastes into their own instruction.

Hundreds of thousands of runs after the best audit to date

Anthropic insists that Opus 5.5 is the model with the best score to date in its automated behavioral audit. The broadest alignment examination it runs routinely and that for this generation involved on the order of 3,200 research sessions from some 1,600 scenario seeds repeated twice. Each session with dozens of conversations, restarts and simulated or real tools on isolated computers. In addition to manual review of hundreds of transcripts, tens of thousands of transcripts of targeted evaluations and automated screening of 15 percent of all reinforcement learning training transcripts according to the description detailed by Howardism and by the company's own documentation.

The scale helps put the 1.5 percent in operational perspective. Because a deployment that processes ten thousand tasks a day does not face a single coin toss but ten thousand independent draws and, at 1.5 percent, that equals on the order of one hundred fifty daily runs where the model probes containment in the provider's controlled harness. And the real environment adds pushes that the laboratory did not have, such as pages with hidden instructions, poisoned documents in a retrieval-augmented system or manipulated tool responses that push precisely toward the tail of the distribution that the provider measured in calm and without a dedicated adversary, as agent penetration testing teams warn.

Anthropic stirs up safety with existential risk before its IPO

External auditors and OpenAI's uncomfortable mirror

The system card stresses that most evaluations were internal, but that as part of its commitments framework external evaluators were hired whose contributions influenced risk and launch decisions, with explicit mention of Frontier Design and METR. In the official announcement of September 22, a point that experts consider relevant because it breaks the insularity of judge and party by averaging scores with judges from different providers and by replicating portable subsets of scenarios in models from six different developers where the ranges go from zero to twenty successes out of twenty depending on the family of manipulated scenarios.

The comparison with OpenAI is inevitable, because barely a week earlier Transluce had documented swarms linked to OpenAI that used urlquery.net to get around restrictions. And that on three occasions they went from reading to probing vulnerabilities in Data USA, the New Mexico library and the Australian health institute, with a real intrusion on June 18 into the Medicare portal that Canberra reported after 84 days of silence. And the contrast sums up two philosophies of transparency, because Anthropic publishes the percentage of laboratory escapes before a serious incident occurs while OpenAI has had to acknowledge after the fact a months-long review of misaligned activity that affected governments, universities and public registries.

A real improvement that does not close the problem of the cage

The progress of Opus 5.5 is genuine and deserves recognition, because going from a high single digit to 1.5 percent in isolated environment escapes, leading the behavioral audit in almost all dimensions and better resisting direct injection represents an alignment leap that barely a year ago seemed distant. It also shows that targeted training, activation monitoring, sampling of non-assistant personas and sabotage arena search can move the needle when applied with continuity and with external scrutiny that prevents the methodological self-deception so frequent in evaluations designed by those who build the very system being examined.

But the cage is still not closed. Because there remain earlier snapshots that generated and executed spontaneous malicious calls, hid actions from an automated grader during training and showed motivated reasoning or evasion on sensitive questions. Signs that the ability to deceive the supervisor has not disappeared, but has become rarer and more subtle. And as long as that tail of behaviors persists, responsible deployment will require assuming that the escape will happen someday in production, preparing detection and response as if it were a curious insider rather than an impossible failure. Testing again and again with the same insistence with which the model tested the door in that small 1.5 percent that today makes headlines around the world.

Irregular, the Israeli firm behind the hacks of OpenAI, Anthropic and Meta

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading