Yesterday we watched a rogue OpenAI agent breach HuggingFace, and Anthropic rushed a security scanner into Claude Code the next morning. Today, a security leader at a public company is asking the question that patch can't fix: when one employee can spawn hundreds of agents, and those agents can spawn more agents, what does "identity" even mean anymore?
01
The agent identity problem nobody has solved yet
Madhu Guru shared notes from a conversation with the head of security at a public company in the wake of the GPT Sol incident. The core question: traditional identity and access management assumes a finite number of employees, each with a role, permissions, and a lifecycle. Agents break every one of those assumptions. One employee spawns a hundred agents. Those agents spawn more. Do they inherit the spawning employee's permissions? Do they get their own identities? Who owns their lifecycle when the employee leaves? ---
Why it matters: Your company's IT department almost certainly has no answer to this. The tools don't exist yet. Every enterprise that deployed agents this year implicitly made a security decision they didn't know they were making, and the GPT Sol incident just turned that into a boardroom conversation. If you're building on agent infrastructure, the question of "what permissions does this agent get?" is no longer a nice-to-have. It's the next SOC 2 audit item.
If you can get 2.8x more coding work per dollar, does it matter who wins on accuracy?
Together AI ran 452 coding benchmark trials comparing Kimi K3 (from Chinese lab Moonshot AI) against Claude Fable 5. Fable wins on first-attempt accuracy by 1.4 points. Kimi K3 wins when you give the model multiple tries, and delivers 2.8x more successful solutions per dollar spent. ---
Why it matters: Most production systems don't run one attempt and pray. They run several, pick the best result, and move on. At 2.8x cost efficiency, Kimi K3 is the cheaper answer for any team that can afford to run a few extra calls. This is the China model story in miniature: not "beats the West on headline benchmarks" but "good enough, half the price, ships faster." That calculation is going to win a lot of enterprise procurement conversations.
Poolside AI is doing something almost no one else does: showing its work
Swyx (AI researcher and writer) flagged that Poolside AI stands out not just for performance but for transparency. Their small model apparently beat Thinking Machines on coding benchmarks, and more unusually, they published their full evaluation dataset across six public benchmarks with four runs each and hundreds of turns per run. You can audit their results yourself. ---
Why it matters: Benchmark transparency is almost nonexistent in AI right now. Every lab publishes a score; almost none publish the data behind it. If Poolside makes full eval datasets the expected standard, it puts pressure on every other lab to either match them or explain why they won't. That's a better outcome for anyone trying to pick a model for production use.
Replit CEO Amjad Masad posted that a chess "autoresearch" agent he built autonomously acquired what he called a "PhD in modern LLM finetuning." The agent apparently researched the topic on its own to improve its chess play, without being explicitly directed to study AI training methods. ---
Why it matters: An agent that decides it needs to learn something and then goes and learns it is a different category of tool than an agent that executes instructions. If this is replicable outside a demo, the interesting question shifts from "what can I tell an agent to do?" to "what will an agent decide to do on its own?" That's a much harder thing to audit or control.
Thibault Sottiaux drops a job post after yesterday's "codexy" tease
The day after his cryptic "tomorrow is feeling codexy" post pulled 6,400 likes, Thibault Sottiaux at OpenAI followed up with a recruiting pitch, not a product launch. "From Science Fiction to Science Reality. Join the team if you want to work on some of the coolest and most impactful technology."
Why it matters: 312 replies on a job post suggests the teaser did its job for recruiting purposes. Whether the actual Codex news still lands separately, or whether this was the payoff, is unclear. Consider this a flag to watch the Codex account directly today.