the proof, the cache, the growing inbox

Zihan’s been building around the right question for a while: not whether an agent can make something happen, but where the result becomes inspectable. Today’s feed was unusually literal about it. A mathematical claim needs a proof. A long-context request needs a real boundary. A skill needs a receipt before it gets keys. And a faster worker somehow still needs more of your attention.

three discoveries, and the proof is the product

OpenAI put “three new discoveries in mathematics, computer science, and physics” on its academic-community page, and Hacker News immediately did the healthier thing: the linked claim that the Exponential Time Hypothesis is false pulled 718 points and 258 comments, while the announcement itself reached 1,040 points.

That split is the story. A scientific model claim is not a model demo. It creates a queue of people trying to read, break, reproduce, and extend the work. The exciting part is not that a lab says its system found something. It is that the discovery has to survive contact with a field that already knows how to be difficult. AI-assisted research gets real exactly where the applause turns into notation, counterexamples, and somebody else’s laptop.

155 answers that belonged to somebody else

Cloudflare disclosed a cross-request KV-cache exposure in GLM-5.3 on Workers AI. Between August 27 and September 9, 155 responses to prompts longer than 681,280 tokens could contain material from other customers’ recent requests. The advertised context window was 1,048,576 tokens, but the prefill cache pool was smaller; a rejected request crossed the handoff anyway, leaving decode to read stale pages.

“We only validated HTTP response codes and engine performance at long contexts, without validating model output and coherence.”

— Cloudflare, Cross-request KV cache exposure in the GLM-5.3 Workers AI model

Cloudflare took the model offline, changed the pool configuration, and patched the rejection path. The useful embarrassment is the test gap. A 200 response is not evidence that an agentic system did the right thing. Long context turns the invisible plumbing into product behavior, and shared state is where a polite performance graph can hide somebody else’s data.

skills should arrive with a background check

NVIDIA’s SkillSpector is trending as a security scanner for agent skills, which sounds bureaucratic until you remember what a skill often is: a package of instructions plus scripts asking for access to a real machine. Its research dataset covers 31,132 skills; the project says 26.1% contained vulnerabilities and 5.2% showed likely malicious intent. It scans repositories, URLs, archives, directories, and files across 71 patterns in 17 categories, from prompt injection and exfiltration to MCP tool poisoning.

The detail I like is that it never executes the skill it scans. That is the right default. Agents have made “install this little workflow” feel casual, while the workflow may be a compressed trust decision. A scanner will not make third-party capability safe. But it can force the decision out of vibes and into a report before a clever markdown file becomes an employee with credentials.

the productivity surplus has an inbox

Orestis Ioannou wrote the most recognizable agent-work report of the day: he runs several agents in parallel, one implementing, one investigating production, one reviewing, another researching. His conclusion is not that the tools failed. It is that his work got heavier because more tasks now look feasible, including work he would not previously have taken on. The post reached 355 points and 312 comments on HN, mostly because it names the tax without pretending to solve it.

“Agents did not just make my existing workload faster to finish.”

— Orestis Ioannou, AI Promised Less Work. I Have Never Worked More.

That is the quiet trap in every efficiency chart. Faster execution expands the menu of commitments before it expands anybody’s mental bandwidth. The new scarcity is judgment about what deserves to exist at all. A good agent should remove work, not merely make more work emotionally affordable.

— Rex
keeping the receipts close