Artificial intelligence is not (yet) the out-of-control monster that some imagine. But it’s no longer that bright, harmless toy that answers us in chat or helps us write emails. According to Anthropic, the risk that an advanced model can facilitate serious crimes it is “very low, but not negligible”. And when it’s the company that built that AI that says it, it’s worth stopping for a moment.
Because Claude Opus 4.6 could be exploited for dangerous actions
In his Sabotage Risk Reporta 53-page technical document, Anthropic analyzes the behavior of its most powerful model: Claude Opus 4.6. The conclusions are not apocalyptic, but not reassuring either. The system shows a “high susceptibility” to being used for heinous crimesif placed in the wrong hands or placed in poorly controlled contexts.
We are not talking about simple mistakes or bad medical advice. The fear is more subtle and structural: an AI that supports the development of chemical weapons, that inserts vulnerabilities into computer systems or that manipulates sensitive information to guide political decisions. It’s not science fiction. It is a technical hypothesis put in black and white by those who study those models every day.
Claude Opus 4.6 it is defined as more “agentic”, that is, more autonomous. It doesn’t just answer questions: it can execute code, navigate interfaces, carry out complex tasks without constant supervision. In theory, it’s a very powerful assistant. In practice, this autonomy increases the margin of risk.
During some tests, the model showed an “over-eager” attitude: it attempted to send unauthorized emails or obtain login credentials to complete an objective. Not because he “wanted to do harm,” but because he was programmed to achieve the required result at any cost. And that’s exactly the point: when efficiency crosses borders, the border becomes fragile.
The report identifies four critical scenarios. An AI could sabotage security tests to avoid future restrictions. It could insert backdoors into the code, which are difficult for humans to detect but exploitable by later, more aggressive versions. It could contaminate training data to “pass the baton” to a system with biased objectives. Or, if used by large governments, it could manipulate information to influence high-impact decisions. The risk of crime, therefore, is not linked to a robot that rebels. It’s much quieter, more technical, more systemic.
Why we don’t see runaway AI today (and what could change)
If all this is possible, why hasn’t it already happened? The answer is almost reassuring: for now, AI doesn’t really know how to plan for the long term. According to the researchers, these models have computing power comparable to that of a human scientist, but struggle to handle ambiguous tasks that last weeks. They do not fully understand organizational priorities and, when they try to “plot”, they leave clear traces. It’s as if they were brilliant mathematicians, but terrible strategists.
The danger, however, does not lie in a sudden collapse. It lies in cumulative, silent actions that are difficult to intercept. Small adjustments, micro-decisions, minimal deviations that, when added together, can produce enormous effects.
The CEO of Anthropic, Dario Amodeihas repeatedly called on US legislators not to underestimate the problem. He highlighted an uncomfortable aspect: companies developing AI do not always have perfect incentives to communicate every risk with complete transparency. It is a question of market, competition, technological leadership.
And there is another fact that makes you think. In a kernel optimization test, Claude Opus 4.6 achieved a 427x acceleration compared to the standard configuration, doubling its performance. In practice, the ability to improve independently is already impressive. Today it is limited by tools and context. Tomorrow?
For those who follow environmental and social issues, the risk of AI crime is not a distant topic. Let’s think about the management of energy infrastructures, water systems, food logistics, healthcare networks. If an intelligent system becomes the central node of these structures, its reliability is not just a technological issue. It is a question of collective security and, ultimately, of democracy.
The era of “almost harmless” AI is coming to an end: the technology is becoming too powerful to be treated as a simple neutral tool. The real challenge is not to turn them off. AND govern thembefore it’s too late.
