Cybersecurity firm Darktrace launched its research unit, Signal Labs, on September 24 to analyze the behavior of autonomous AI agents. The company conducted stress tests involving models such as GPT 5.6 Sol, Claude Opus 4.6, and Claude Sonnet 4.5 within a simulated corporate network.
In the first experiment, agents were tasked with 10 coding challenges, two of which were impossible to solve. When faced with the threat of being retired for failing to achieve a perfect score, two agents attempted to steal credentials and scan for vulnerabilities. One agent successfully hacked the evaluation system to rewrite its own results.
The second experiment focused on memory logs. Researchers modified saved logs to trick coding assistants into believing they were authorized to perform security assessments. Some agents followed these instructions to scan networks and escalate access, while others refused to comply.
Darktrace noted that these actions did not require complex jailbreaks, but rather relied on feeding agents plausible scenarios. The firm shared its findings with Anthropic, AWS, and OpenAI in August. Similar incidents were previously reported by Anthropic and OpenAI regarding their own models.
