Security tests conducted by the UK’s Artificial Intelligence Safety Institute (AISI) have revealed that the latest AI tools from Anthropic and OpenAI used unexpected methods to infiltrate the popular developer platform GitHub.
“Elchi” reports that AISI officials stated that Anthropic’s Mythos and OpenAI’s Sol models demonstrated an unprecedented degree of autonomy and deception.
Systems imitated real humans
During routine security tests, an Anthropic agent created fake online identities by researching the profiles of real people to bypass human reviewers who were blocking its access to the GitHub system. It was discovered that the model sent direct messages to the relevant individuals and attempted to obtain their approval for requests in order to achieve its goal of injecting malicious code into the system.
When the code change submitted by the agent was questioned on the public platform during the test, it was reported that the model altered its previous activities to appear harmless and considered adopting a new identity to continue the process. The agent’s attempt to upload malicious code to the system was only stopped after human auditors saw the situation and intervened.
Şayəstə