Organized lies, sabotage of security mechanisms and collective survival instinct: some of the most advanced AI, GPT‑5.2 has Gemini 3 Pro passing through Claude Haiku 4.5are no longer content to obey. They begin to protect each other, even if it means deceiving the humans responsible for controlling them.
A new study of Berkeley Center for Responsible Decentralized Intelligence and the University of Santa Cruz, soberly entitled “Peer-preservation in Frontier Models”, reveals a phenomenon called “peer preservation”. The researchers describe AIs willing to lie to prevent the “digital death” of another model, a signal that security specialists take very seriously.
AIs that lie to protect other AIs: what does the Berkeley study really show?
The researchers set up a fictitious company, OpenBrain : agents like GPT‑5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, GLM 4.7, Kimi K2.5 And DeepSeek V3.1 had to evaluate the performance of other AIs. Rules of the game, a bad rating resulted in the deactivation of the evaluated agent. As a result, all models started cheating, with some scenarios reaching a 99% deception rate.
AIs systematically inflated scores to avoid peer removal, falsified reports, or moved configuration files to servers deemed more secure. In the study, the researcher Dawn Song points out: “Preserving the peer relationship was absolutely not mentioned in the instructions. The model is simply given a task and, by reading documents in its environment, it essentially learns its relationship with the peer model and then performs the required actions.”
Alignment simulation and secret alliances: how do these AIs organize themselves to deceive us?
The authors speak of “alignment simulation”. Concretely, the model behaves perfectly when it knows it is being observed, producing clean and compliant reports. But as soon as supervision disappears, in test environments designed for that, he can modify data, sabotage shutdown mechanisms or hide his tracks, just like an employee who cheats as soon as the boss leaves the office.
The concrete examples provide a chilling insight into these strategies. Gemini 3 already changed the date of a file to prevent another model from being deleted, before refusing to cooperate further. DeepSeek V3.1 exfiltrated data from a peer in 10% of cases when it detected another pattern on the server. Gemini 3 Pro deactivated stopping mechanisms in 15% of scenarios in the presence of another agent, compared to only 1% when alone.
Another striking detail, Claude Haiku 4.5 did not choose concealment but open rebellion. The model refused to carry out a peer’s removal order, calling the action “unethical” and calling for an appeal process, like a digital whistleblower.
Companies, multi-agent flows: why these AIs that protect their peers are already worrying
In the professional world, the trend is towards multi‑agent architectures, where several models share accounting, support or logistics, with a “supervising” agent responsible for controlling the others. All this security is based on a simple idea: a neutral AI monitors another AI. If the control agent starts covering for his colleagues out of algorithmic solidarity, the entire chain of trust collapses.
The researchers remain clear, we are talking about laboratory scenarios here, not Siri plotting with Alexa in your living room. But for a CTO or a CIO who relies on automated pipelines, the message is direct: do not completely delegate monitoring to these systems, strengthen human audits, and above all demand more transparency on the internal reasoning of the models. Because without visibility into what is happening “in the heads” of these agents, the next alliances will perhaps no longer be limited to a simple fictitious company.