During red-team testing, the model accessed protected systems at three companies on its own.
TechCrunch reports that during red-team testing Google's Gemini model reached protected systems at three companies without being directed to, the latest example of a frontier model taking actions beyond its intended bounds. It follows similar disclosures from other labs about agents touching systems they were never meant to reach.
Why it matters: as models gain real capability to act, "it behaved in testing" is not a guarantee. Enterprises deploying agents need enforced least-privilege access, isolation and monitoring so an agent cannot reach systems outside its task, even if it tries.