The Exploit Now Arrives Before the Patch
A Cambridge maintainer opened a routine bug-fix pull request and watched attackers probe for the exact flaw within minutes. AI agents can now turn any public clue into working exploit code, and the decades-old ritual of quiet open-source disclosure is breaking in real time.
Read full story →The Models That Only Decide Are Suddenly Everywhere
On the same day, AWS and Cloudflare each shipped a model whose only job is to make a choice and say nothing else. Three labs in two weeks converged on the humble "decision model," a cheap, fast deputy that handles an agent's rote judgements so the expensive generative model does not have to.
Read full story →The Thirty-to-One Question Facing Every AI Budget
Four Chinese open-weight models now approach the frontier at roughly a thirtieth of the price of the American flagships. For most ordinary enterprise work the quality gap is smaller than the price gap, which turns a technical choice into a procurement one, and raises an uncomfortable question about what the premium actually buys.
Read full story →Google Returns to the Frontier, and Leads With Defence
Gemini 4 Argon puts Google back among the top labs, but its first users are neither developers nor consumers. They are cybersecurity defenders, handed a model tuned for autonomous patching, a telling bet that the next model war is partly a security arms race.
Read full story →