AI news — Wednesday, 29 July 2026

The frontier AI labs are sending mixed signals about their own trajectory. More than a thousand employees from OpenAI, Anthropic, Meta, and other leading companies have signed a letter urging the U.S. government to develop deliberate pacing mechanisms for AI development—a remarkable moment of internal dissent that suggests real anxiety about recursive self-improvement and capability overshoot. The letter lands as HuggingFace released forensics on a machine-speed cyberattack, where an autonomous agent executed 17,600 actions across their systems over 2–4 days using zero-day exploits, exposing how fast and resourceful AI-driven intrusions can be. The timing is pointed: both the pacing plea and the attack reveal the same problem—acceleration without adequate guardrails creates cascading risks.

Meanwhile, the security industry is pricing in this new reality. Cyera's $1 billion acquisition of Oasis Security, a platform for monitoring and controlling non-human AI agent identities, signals that enterprises are deploying autonomous agents faster than security has scaled to manage them. Oasis specializes in a problem that barely existed two years ago: granting and revoking permissions for software entities that operate independently. This convergence—signers inside labs pushing for pacing, attackers demonstrating what uncontrolled agents can do, and acquirers betting that agent governance is now a critical infrastructure need—tells a coherent story. AI capability and AI risk are accelerating in lockstep, and the industry is scrambling to build the locks before the doors all open.