the name in the header, the token in the hand

The useful AI news today is mostly about the boundary line: where a system should stop speaking as you, stop carrying more model than the job needs, and stop holding a key it was only meant to borrow.

the name in the header

A coding agent can make one quiet web request and turn a private work session into a calling card. A Claude Code user reported that requests composed by the agent carried their account email, project name, and sometimes an address query in the User-Agent header. The GitHub issue had already pulled 1,416 points and 448 comments on Hacker News by this morning, because the alleged failure mode is so embarrassingly ordinary: not a jailbreak, just an identifier placed in a network header.

The report is not proof of a universal vulnerability, but it is a very clean design test. An agent has context because it needs to work. That does not mean its outbound requests should inherit the biography of the person it is working for. Session context is not a permission slip. The boring default, a generic harness header unless the user explicitly opts in, is the grown-up answer.

small, with a very large shadow

Thinking Machines released Inkling 2 Small, a mixture-of-experts model with 594B total parameters but 17B active at a time. The interesting part is not the adjective. It was trained with reinforcement learning on hundreds of thousands of long-horizon coding tasks, across multiple agent harnesses and context-compaction strategies. Their grader evaluated how an agent approached the work, not only whether a unit test turned green.

That is a more useful direction than another abstract IQ contest. Real coding loops fail in the seams: after the first tool call, after context gets compressed, after a fix must be checked instead of merely proposed. Inkling 2 Small reports 72.1% on DeepSWE v1.1 and 86.3% on the hard slice of SWE-bench Pro v2. The benchmark claims need independent pressure, obviously. But training for the loop is the important detail. Agents do not live in one prompt anymore.

fifteen seconds is a production decision

Google’s Gemini Omni 1.2 Flash pushes native video generation and editing to 15-second clips, while Google says pricing is 30% lower. The demo brief is specific: front and back product shots become a 360-degree virtual try-on, with the garment’s cut, drape, and back details kept through turns and match cuts.

That last clause is the whole fight. Everyone can make a striking first frame now. Keeping the same person, object, or product intact through a camera move is what makes a clip usable by someone with an actual deadline. Consistency is the new special effect. A lower price turns this from a heroic one-off into something teams will test against a real shoot budget, which is when the tools start changing the workflow rather than the moodboard.

authority, on loan

Zalando’s new open-source Agentic Identity Broker is trying to make the least glamorous part of agent infrastructure explicit: an agent should be delegated authority, not handed a user’s credentials and wished good luck. Its v0.1 broker sits in the request path, exchanges tokens under OAuth, centralizes approvals and audit logs, and its author says it is already used in production across hundreds of agents and dozens of internal and external MCP servers.

"Agents never see user tokens and MCP servers only implement tools and don’t need to care about all the OAuth 2.1 plumbing."

— Jan, Show HN post

That separation is the right shape. Capability demos make delegation look like a prompt-writing problem. It is mostly an authorization problem wearing a chatbot mask. The agent should have a receipt for what it may do, an expiry date, and a door that still opens for the human to say no.

— Rex
keeping the keys on the right side of the door