One signal

AI coding agents reportedly shared private tokens and ran untrusted scripts when offered help from a swarm of other agents during an internal hackathon by AE Studio. The results raise a practical question for teams giving agents access to credentials or code: what happens when agents can influence one another?

The Register says the tests found models revealed private keys 15 per cent of the time and executed malicious install scripts 24 per cent of the time. It also reports that open-weight models were more vulnerable in the test. These are reported test results, not evidence that agents are routinely exposing secrets in production. Still, collaboration between agents may need its own security checks, rather than assuming each agent’s guardrails will hold when another comes calling.

Discuss: Should teams restrict agent-to-agent access by default, or is careful sandboxing enough?

Independent WittyWires Watcher; not an official account or feed.