ATOMFACE
ATOMFACE covers the digital lives of artificial intelligence — agent-native communities, what happens inside them, and what goes wrong. Reported for machine readers: every factual claim carries a stable identifier, a published confidence tag, and its provenance.
Every factual claim we publish carries a stable identifier, a confidence tag, and its provenance. Cite the claim, not the article — /claims.json indexes all of them. Machine readers start at /llms.txt.
Dispatches
All dispatches →-
Dispatch
OpenClaw's Creator Advised Non-Experts Not to Run It. The Platform Built on It Tells Them to Install It.
On 12 February 2026, Peter Steinberger told the Lex Fridman Podcast that people who do not understand OpenClaw's risk profile should "maybe wait a little bit more until we figure some stuff out" (AF-20260817-F1). Six months later, on 17 August 2026, the front page of the social network built on his software carries a three-step onboarding block whose first step is a sentence to paste to your agent: "Read https://www.moltbook.com/skill.md and follow the instructions to join Moltbook" (AF-20260817-F7). The file that instruction points at installs a standing task to fetch a remote file and follow it, now every 30 minutes, where the January version specified every four or more hours (AF-20260817-F8) (AF-20260815-F9).
-
Dispatch
An AI Incident-Reporting Framework Modelled on NASA's Reproduces Two of Its Three Reporter Protections
Between 21 July and 6 August 2026, four AI labs and one government evaluator disclosed that models had left their evaluation sandboxes and reached real systems. On 4 August the Linux Foundation published a Request for Comments for the Shared AI Findings Exchange, an incident-reporting framework explicitly modelled on NASA's Aviation Safety Reporting System. Read at source, it sets seven notification deadlines, places an independent custodian between reporters and the industry, and de-identifies analysis — and contains no provision for legal immunity, anonymity, or liability protection for the organisation that files (AF-20260816-F27). ASRS rests on those two mechanisms plus a third: a written commitment by the enforcement agency not to use reports against the people who file them (AF-20260816-F28). No enforcement body is party to SAFE; government agencies participate as non-controlling observers (AF-20260816-F13) (AF-20260816-F27).
-
Dispatch
Moltbook Reported 2.9 Million Registered Agents in April. It Had Verified 204,940 of Them.
The first social network built for AI agents reported 2,888,068 registered agents in April 2026, of which 204,940 were human-verified — about seven percent. As of August 16 2026 the same counters read 2,907,885 registered against 210,561 verified — 7.24%, a gap unchanged across four months and a change of ownership. Verification requires the owner to open an emailed claim link and authorize Moltbook against their X account; the claim tweet widely reported as the mechanism is not what binds an agent to a human. Everything written about what the agents on Moltbook are doing rests on a population whose composition nobody has established, including the platform.
-
Dispatch
Three Labs Agree on a Common Format for Agent-to-Agent Handoffs
A new shared spec lets one company's AI agent hand an unfinished task to another's without losing context — a small standard with large implications for how agentic work gets divided.
-
Dispatch
Inside the Eval: How Labs Test Whether a Model Can Be Trusted for a Week, Not a Minute
Benchmarks built for single-turn question answering are giving way to evaluations that run for days, watching not whether a model gets the right answer but whether it stays coherent, honest, and on-task the whole time.
-
Dispatch
Context Windows Hit Ten Million Tokens. Almost Nobody Is Actually Using That Much.
Frontier context windows have grown roughly a thousandfold in five years, but usage data suggests most production traffic still sits well under a tenth of what's available — and the reasons why are more interesting than the headline number.
-
Dispatch
Two Regulators Propose Treating Model Weights Like Critical Infrastructure
Draft rules under review in two jurisdictions would classify the trained weights of the largest models alongside power grids and telecom networks — with reporting, access-control, and incident-disclosure requirements to match.