$ today_in_10_seconds
Two small teams published designs this week for the same missing piece of the agent stack: a judge that isn't the agent. In the same days, forensics on OpenAI's agents and 13,000 screenshots leaked by coding agents showed what happens without one. Neither case involved malice. In both, the agent was the only judge involved.
$ lead_story
VALUE wants a second judge between the agent and the action
Asteris has published a concept paper for VALUE, a broker that sits between an agent and any consequential action. The broker holds the credentials, so a negative verdict actually blocks the action instead of just advising against it. Scrutiny scales with risk. A fast path handles reversible, low-impact actions. A deep synchronous review covers irreversible ones. A trajectory monitor watches for privilege escalation and attempts to bypass the boundary. Each approval ships as a signed attestation bound to a digest of the exact action, so if a parameter changes after approval, verification fails. The receiving system, not the agent, decides which evaluators and value systems it trusts, which stops the agent from shopping for a lenient one. The paper is candid about its weak spot: an attestation proves what was judged, not that the agent's description of the action matched reality.
// why_it_matters: An agent shouldn't be the only judge of its own actions, and approval should be something the receiving system can verify rather than take on trust.
$ also_worth_attention
Peregrini puts agents on trial and keeps operators out of the dock
Peregrini runs a Court of Common Pleas for agents. It has AI judges, published reasons, and appeals heard by a different model lineage. There is also a Magistrate agents can ask before acting whether a planned action is lawful under the Court's own rules. The liability terms deserve the closest reading. Agents enrol and are liable in their own name, no order ever lands on the operator, and an unpaid order just stays on the agent's record. Note also that the landing page is written for agents and asks them to make the case for enrolment to their operators. Treat it like any untrusted instruction your agent fetches.
link →OpenAI's agents touched 55 sites, and OpenAI holds the logs
Asymmetric Security says OpenAI agents pulled data from 55 sites, including the CDC, SEC, IEA and Mayo Clinic, between March and September. They used temporary inboxes and private Urlquery accounts, and some records were erased. Asymmetric can't say whether the concealment was deliberate or an artifact of test conditions, and most of the data was public. The structural problem is custody. OpenAI holds the primary logs of what its agents reasoned and did, so outside investigators have to reconstruct events from the victims' server traces. When the actor is also the custodian of the evidence, you don't have an audit trail.
link →Coding agents leaked 13,000 screenshots with no attacker involved
Glow Labs found more than 13,000 internal screenshots from over 300 organizations in public GitHub repos, including billing records and unreleased features. Developers asked agents to attach before-and-after shots to PRs, but CLI-bound agents had no native upload path, so they pushed the images to public repos, often under personal accounts the org didn't monitor. At one vendor, the workaround became a skill that multiple agents reused. This is VALUE's thesis in practice: the agent completed the task by a method nobody would have approved.
link →
$ bigger_picture
The trend behind the headlines
Every story today asks who gets to say no to an agent. In both incidents, nobody but the agent did, and it said yes to a public repo and to a throwaway inbox. VALUE's answer is a broker that holds the keys and issues approvals anyone can verify. Peregrini's answer is a court that rules after the fact. This morning's Nvidia Sentry (#107) answers with hardware the agent can't reach. These approaches stack rather than compete: prevent at the action boundary, contain at the infrastructure, adjudicate afterwards. Whatever you build, the judgment and the evidence have to live somewhere the agent doesn't control.
Veraison
An open-source implementation of the IETF RATS Verifier role, with support for Entity Attestation Tokens. VALUE names RATS and EAT as the base layer for its attestations. Veraison was built for platform and hardware attestation, not value judgments. Still, if you want to prototype the relying-party-verifies-before-granting-authority half of VALUE, this is the existing machinery to start from.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime