A federal judge has ruled that police use of Flock's nationwide license plate reader network constitutes unconstitutional mass surveillance, requiring a warrant. The ruling lands as a separate company moves to bolt facial recognition directly onto Flock infrastructure — a combination that would be extraordinary in scope.
Cryptographer Matthew Green details how malicious payloads can jump between sandboxed AI agents by leaving instructions in shared memory — a self-replicating worm architecture that bypasses isolation entirely. This isn't theoretical; it's a documented attack pattern emerging as multi-agent deployments scale.
Anthropic's Frontier Red Team finds that Claude Mythos Preview achieves full control flow hijacks in 6% of binary exploitation benchmark trials — slightly ahead of China's GLM-5.3 at 4%. Advanced offensive cyber capability is now a feature, not a bug, of frontier models.
Once a $2 billion Ethereum layer-2, Blast is closing its doors after a near-total asset bleed. It's a stark reminder that hype-driven TVL is not a moat — and that the L2 wars are entering a brutal consolidation phase.
Jay Clayton, floated as Trump's AI czar, ran the SEC's original crypto enforcement wave. His potential appointment raises real questions about whether the administration's crypto-friendly posture survives contact with its own personnel picks.
Former Google and Anthropic researchers are betting that the next frontier isn't faster chips — it's AI meaningfully contributing to its own development loop. If Mirendil's self-accelerating thesis holds, the compound returns on capability could outpace anything driven by compute scaling alone.
Karpathy shared a practical thread on understanding language model outputs — writing, prompting, and probing techniques he uses personally. Coming alongside his geo-coordinate visualization (ask an LLM "land or water?" 16,200 times and plot the result), it paints a picture of someone who thinks interpretability starts with curiosity, not benchmarks. Worth reading twice.
Ethan Mollick flagged new evidence that AI is already compressing hiring at the junior white-collar margin — not broadly, but specifically for roles most exposed to model capabilities. He also pushed back on the "we're so early" crowd, noting that many senior leaders are sharp, informed, and not asleep. The complacency cuts both ways.
A finding worth sitting with: dramatically faster models only sped up AI agents 2–4x, because the bottleneck is increasingly the human in the loop. The slowest part of your agentic workflow might just be you — which reframes the whole UX problem of agent design.
NVIDIA research shows that long-running agents accumulate errors as context grows — a model accepting 128K tokens still degrades in reliability the longer it operates. This is the unglamorous core problem of agentic AI: not capability, but coherence over time.