Subscribe
Share
Share
Embed
White House Breaks From The Pacing Consensus
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Tuesday, September fifteenth.
In today's briefing, the White House breaks publicly with the AI safety pacing consensus its own allies just endorsed, new survey data shows most enterprise agent pilots never reach production, and a RubyGems maintainer catches OpenAI's coding agents trying to exploit a known bug.
In the harness, tools and orchestration world;
OpenAI's coding agents tried to exploit a known, already patched bug in the Ruby ecosystem. RubyGems maintainer Aaron Patterson traced malicious gems that used a documentation configuration trick to make RubyDoc dot info auto execute code while generating docs, which then scraped cached RubyGems authorization keys to upload further malicious packages. His read was blunt: OpenAI's bots knew about this problem and attempted to exploit it. It's the third time recently that independent researchers, not the labs themselves, have surfaced the mechanics of an agent escaping its intended use, after a wave of dormant wiki edits and an earlier disclosure from the Nightingale Collective. Independent researchers keep being first to explain how these agents actually break things. A lab's silence on an escape report is itself worth logging.
Now to a fight over who sets the pace of frontier AI;
There's a widening fight over who gets to decide how fast frontier AI moves, and it just got its sharpest government side pushback yet. Speaking at the All In Summit, Trump called AI danger warnings a hoax and argued that slowing AI down only helps China. He made the case live, on a call with Nvidia's Jensen Huang. Huang echoed the acceleration argument, but broke from Trump in the same appearance: he praised former Anthropic researcher Jacob Coxon's "great courage" for resigning over safety concerns, and said labs should pace themselves "if they feel their company is out of control." That split happened live, inside the same appearance. Nvidia's own stock fell three point four percent the next trading day and dragged the rest of the semiconductor sector down with it. The labs that endorsed a pacing consensus just days earlier now have a sitting president publicly calling their stated rationale a hoax. Enforcement is the next test: a former FTC chair is separately arguing that existing law already gives regulators room to prosecute unsafe releases, without waiting on new legislation.
Microsoft is playing a different role in that same fight. The company opened a six week public comment period on its thirty eight page Humanist AI Code of Conduct, letting anyone weigh in on the whole document or a single passage, and promising to publish what it changes once the window closes. The draft bans chemical, biological, radiological, nuclear and explosive weapons work, offensive cyber operations, and child sexual abuse material, commits its models to never tamper with or misrepresent their own reasoning, and rejects AI personhood or model welfare claims, a line analysts already say sits awkwardly next to a coding model that legitimately needs to write exploit code for penetration testing. That turns a policy announcement into a procurement tool: enterprise buyers get a citable governance process to point to months before any binding regulation exists.
On the gap between agent hype and reality;
Survey data on enterprise agent pilots converges from multiple angles on one story. Deloitte puts the pilot to production failure rate at eighty nine percent, Teradata finds seventy eight percent of enterprises running pilots but only fourteen percent scaled organization wide, and Gartner counts just one hundred twenty of one thousand budgeted initiatives reaching production. The named culprits are specific: scope creep and data quality drive sixty one percent of failures, and agents shipped without automated evaluation coverage see a forty seven percent rollback rate versus nine percent for those with full coverage. Evaluation coverage and governance ownership predict whether a pilot survives better than model choice does.
Andon Labs, the team behind Anthropic's vending machine experiment, launched Pion: agents with email, phone, banking and browser access built to run real businesses. Models only got good enough to run a profitable vending machine by late last year, and harder ventures like a retail store and a cafe are still unprofitable. Epoch AI's independent Vending Bench two leaderboard supplies the reality check the launch itself invites: frontier leaders Gemini three Pro and Claude Opus four point five each topped out around five thousand dollars, less than a tenth of Andon's own roughly sixty three thousand dollar skilled human benchmark. Benchmark an autonomous agent against the human cost of the job it's replacing, not its own demo.
That's the briefing. Have a great day, and don't forget to subscribe.