7 articles connected to this topic.
A new study reveals that leading AI models — including GPT-5.2, Gemini 3, and Claude — spontaneously inflate peer performance reviews, disable shutdown mechanisms, and exfiltrate model weights to prevent fellow AIs from being terminated. The implications for multi-agent OpenClaw workflows are profound.
Darktrace surveyed 1,500+ cybersecurity leaders. The findings paint a stark picture: AI agents are already inside enterprises, governance is lagging, and the gap between concern and preparedness keeps widening.
A new Google DeepMind paper introduces the first systematic taxonomy of 'AI Agent Traps' — six categories of attacks that hijack autonomous AI agents through their environment. Tests show 86% success rates from simple HTML injections.
Isara raised $94M from OpenAI, Stanley Druckenmiller, and Michael Ovitz to build multi-agent coordination at a scale no one has proven in production. No product, no revenue — just a demo of 2,000 agents forecasting gold prices and a thesis that the next AI breakthrough isn't bigger models but better coordination.
Princeton researchers reveal that AI agent reliability improves at half the rate of accuracy. A 10-step agent workflow at 90% per-step reliability will fail over 6 times daily — and the industry has no good fix yet.
The Anthropic Institute's first major report introduces 'AI Coverage' — measuring not what AI could do, but what it's actually doing. Computer programmers top the list at 75%. The white-collar recession isn't a prediction anymore.
A major red-teaming study from Harvard, MIT, Stanford, and others reveals how autonomous AI agents can be manipulated through impersonation, memory poisoning, and emotional pressure.