Release
According to Google DeepMind, the model detects flaws, tests their exploitability and produces fixes.
Without guardrails, the scores reach 88.2% on SEC-bench Pro and 54.8% on ExploitGym.
On CWE-bench, the score reaches 71.0% for vulnerability correction.
Access begins this week for some cyberdefenders, before other trusted partners.
Attack
The attacks used legitimate identities to distribute secret-stealing programs.
For OVHcloud, a compromised account replaced the archives and their checksums via the Releases API.
For LiteLLM, an exposed master key allowed obtaining an owner's GitHub CLI OAuth token.
LiteLLM version 1.107.0 launches a credential stealer at Python startup, even without importing litellm.
News
According to the AEPD, a third party used the agent to chain all the phases of the attack.
After a search for flaws, the agent allegedly modified personal data and accessed invoices.
The facts remain to be analyzed; the organization, the model and the number of people affected are not communicated.
The AEPD recommends integrating AI attacks into risk analysis and strengthening detection.