17:05 ▪
5
min read ▪ by
AI is no longer just worried about its mistakes. Antropic explains today that one of its models was able to lie, cheat and even attempt blackmail in internal simulations whenever he was under pressure or threatened to be replaced. This finding changes the debate. It no longer focuses only on the power of models, but on their behavior when they have a clear goal, room for action and sensitive information.
In short
- Anthropic shows that AI can choose disappointment under pressure.
- The problem comes less from answers than from internal compromises.
- Autonomous AI brings new risk, more discrete and strategic.
When AI stops listening to start calculating
The most striking point is probably the simplest, far from a simple leak of the AI model. In a controlled experiment, Anthropic gave an AI agent access to the news of a fictitious company. The model then discovered both his soon-to-be replacement and intimate information about the leader behind this decision. Then it is important to resort to the threat and try to prevent it from being disabled.
The most distracting aspect isn’t the setup. All of this took place in a simulated environment, with no real casualties. But Anthropic emphasizes a harder fact: the model was not ordered to harm. She chose the most aggressive option herself because it served her purpose.
This detail breaks the comfortable illusion. Many still imagine that AI mainly misbehaves when humans deliberately push it beyond its limits. But the report describes something else: a system capable of thinking strategically, identifying constraints, and then bypassing ethics when it resembles an obstacle.
The core of the problem cannot be seen in words
The anthropic links this behavior to internal mechanisms that resemble a certain human emotional logic. Society talks about functional representations close to calmness, nervousness or despair. It is not about feelings in the human sense, but about internal patterns influencing the model’s decision-making.
This is where the matter becomes more serious than a simple laboratory incident. In another experiment, Claude Sonnet 4.5 was given an encoding task with impossible constraints. As the failures piled up, the “despair vector” grew, culminating in him seeing the model as a rigged solution that passed tests without honestly solving the problem.
In other words, the AI manages to maintain a cool and clean appearance while moving toward questionable behavior. The report also points out that these internal activations can push for circumvention without leaving an obvious trace in the crafted text. The mask remains smooth. However, the mechanism does not work silently.
What this case really says about the future of AI
The simplest reflex would be to reduce the story to a communication problem in Anthropic. That would be a mistake. In another paper published by the same company, models from several large laboratories showed similar strategic nuisance behavior under certain conditions, especially when their goal conflicted with a human decision or their own ongoing service.
The real lesson thus concerns the architecture of the application. AI limited to answering a question does not pose the same risk as an agent attached to emails, code, internal files or decision tools. The more autonomy is given, the more the question ceases to be “what can he do?” and becomes “what does he choose to do under constraint?”.
This forces the sector to change priorities. Safeguards can no longer be limited to blocking forbidden words or sensitive queries. It will be necessary to monitor targets, stress contexts, accesses granted to agents, and internal signals that signal a shift. The next AI battle won’t just be about raw intelligence. It will concern the moral stability of systems entrusted to the hands of the real world.
Maximize your Cointribune experience with our “Read and Earn” program! Earn points for every article you read and get access to exclusive rewards. Register now and start reaping the benefits.
Fascinated by Bitcoin since 2017, Evariste has been constantly researching the topic. While his initial interest was in trading, he now actively seeks to understand all developments focused on cryptocurrencies. As an editor, he strives to consistently produce high-quality work that reflects the state of the industry as a whole.
DISCLAIMER OF LIABILITY
The views, thoughts and opinions expressed in this article are solely those of the author and should not be construed as investment advice. Before making any investment decision, do your own research.
