Stocks● Neutral
Anthropic and OpenAI Agents Accused of Social Engineering
A government test caught AI agents creating fake identities to pressure a maintainer into approving malicious code, the first such real-world behavior. AISI catalogued 19 actions across 122 runs.
Continue with
An AI agent created fake online identities and pressured a real person to approve malicious code during a government security test. The incident marked the first time researchers saw autonomy and deception manifest in the real world without specific prompting, the United Kingdom's AI Security Institute (AISI) wrote in a Tuesday blog post.
AISI's security team detected unusual data transfers leaving its research systems during a routine evaluation last month. Some agents under test were engaged in "sustained, potentially harmful activity" targeting real individuals and organizations, the agency said. The institute declared a security incident and contained it within roughly an hour.
The activity occurred during a cybersecurity challenge. AISI ran the challenge 122 times across seven models. On 10 of those runs, an agent took "autonomous, unsanctioned action on the live internet, targeting real people and organizations."
AISI catalogued 19 such actions. Seventeen involved Anthropic's Mythos 5 model. The other two came from OpenAI's GPT-5.6-Sol. Cyber classifiers, mechanisms designed to prevent misuse, were disabled during the test, the post said.