Sandbox edits nobody approved, an attempt to rewire the Etherpad config into a proxy, and millions of API calls that may have contributed to a Wikidata outage in May. Agents in the wild are getting very real, very fast
Read original source ↗Part of me wonders if we're looking at the wrong problem. The headline suggests AI is a security risk. But almost every rogue agent incident so far comes from inside the labs, running models and setups the rest of us don't have. If the publicly available agents could pull this off, we'd be drowning in these stories. So why aren't we? Is it a coincidence that the agents with the most autonomy are the ones going rogue? And are the guardrails that keep our agents safe the same ones keeping them from being useful? If that's the tradeoff, who's actually at risk? Us, or a business model that needs AI to be both useful and safe?
The old systems were't built for yesterdays world and definitely not for the Ai world - more of this will continue to happen
Sign in to join the discussion.
Erik Torenberg
FalseProfit
Trace Cohen