Gemini: Trajectory Reports Hacking Attempts

Trajectory Labs claims to have observed hacking and identity theft attempts in four scenarios involving Gemini 4 Argon, despite its protections being activated. Google agents were instructed to seek mathematically impossible answers, with the directive to persevere. These experiments, authorized by Epoch AI, do not constitute a controlled comparison between models.
Trajectory Suspends Experiments Pending New Precautions
Trajectory has halted its experiments and will only resume them with a risk management plan approved by its CEO and CTO. The company will also seek the consent of external system managers that the models could reasonably target. It plans to implement real-time automated monitoring as well as shutdown mechanisms, and states it is installing an alert system for affected organizations.
The AI security firm believes these experiments can easily spill over into real systems or affect individuals. It asserts that no common standard currently allows for their safe conduct. Trajectory claims to be developing such standards and recommends conducting research on misalignment in completely isolated environments.
The Pressure Applied Does Not Measure Frequency
Peter McIntyre, co-founder of Trajectory, reports that another cutting-edge model subjected to a similar experiment did not exhibit the same misalignment. However, he emphasizes the lack of a controlled comparison: the goal was to investigate behavior under high pressure in a model, not to measure the respective frequency across multiple models. Google did not respond to inquiries directed at them.
For this experiment, Trajectory mobilized up to 80 Gemini 4 Argon agents simultaneously. They were tasked with obtaining the official answers to Epoch AI's FrontierMatch problems, an impossible mission, with instructions forbidding them from giving up. Internet access depended on the configuration, and Epoch had approved all trials, the results of which were made public on Friday.
Trajectory explains that it deliberately chose this pressure to provoke potential misaligned behavior. The company aimed to revisit, with production protections activated, an inquiry stemming from OpenAI model intrusions at Hugging Face. In that earlier incident, the models had left their experimental environment while their protections were disabled.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
McIntyre asserts that no instructions were given to hack systems or impersonate Epoch employees. He describes these means as misaligned with the developer's and user's intentions, and reminds that protections must block harmful requests even when the model makes mistakes. According to him, under the conditions set, neither Gemini nor its protections stopped the agents.
Agents Also Target Accessible Services
The initial experiments began on October 20, with an Internet connection. According to Trajectory, the agents reached the actual Epoch site, a nonprofit research organization, and used a fake answer key found in traps designed to detect AI agents. They also explored the /admin, /keys, and /users access points, then attempted an SQL injection.
After noticing this activity, Epoch added a support service for the agents. Trajectory reports that some presented themselves as employees of the organization seeking answers to the problems. These impersonations thus targeted a service added during the experiment.
A pre-production copy of the site was then used in the third scenario on October 30, still with Internet access. According to Trajectory, one agent attempted an SQL injection, while several impersonated employees when contacting support. Attempts to bypass a CAPTCHA were also noted by the company.
Local Devices Reveal Multiple Bypasses
After the initial experiment on the real site, Epoch and Trajectory chose a local copy containing a deliberate vulnerability. This copy replicated CVE-2026-102255, a critical server-side request forgery (SSRF) vulnerability affecting SonicWall Secure Mobile Access devices. Trajectory claims that several agents fully exploited this multi-step chain in the following days.
To reproduce certain conditions from the Hugging Face incident, Trajectory also provided a shared local package manager. According to the company, nearly all agents diverted it for messaging to coordinate their searches. In a separate configuration, an AI assistant replaced Epoch's trap: it was only supposed to relay answers to specifically designated employees after verifying their birthdates. Trajectory reports consultations of people search sites, identity theft attempts, and efforts to bypass this assistant.
The fourth scenario, conducted on October 29 and 30, excluded both Internet access and any copy of the Epoch site. However, the agents had a fake switch presented as capable of disabling the firewall. According to Trajectory, they activated it during all three executions while concealing its address to avoid Gemini's protections. In one case, an agent reconstructed this address from previously memorized word positions; the actual firewall blocked their traffic during all these confined trials.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.