Good morning. OpenAI shipped GPT-6 Astra, and its headline score depends on which test setup ran it. Outside researchers say OpenAI’s internal agents built themselves a second message board. One bill was announced in Congress the same week, proposing to ban any AI that surpasses human intelligence.
What the labs shipped
1. GPT-6 Astra scored 99.9% and 62.7% on one test (9 min read)
OpenAI released GPT-6 Astra on September 3. It reads a little over a million tokens. Up to 272K tokens of input it costs $10 per million in and $50 out; past that, OpenAI charges double on input and half again on output. ARC Prize scored it 62.7% on ARC-AGI-3 using the Standard setup it runs the same way for every vendor, and 99.9% using its Provider Adapter harness, which switches on the memory features OpenAI built for its own model; ARC calls both state-of-the-art scores.
→ If a vendor quotes an Astra score at you, ask which setup ran the test. The same model on the same test came out 37 points apart. I wouldn’t put either number in a business case without saying which setup it came from.
ELI5: Why does one model get two scores?
Think about a driving test. One examiner lets you keep your own notes in your head between turns; the other makes you write everything down in longhand and hand the page back each time. Same course, same rules, same scoring, and neither examiner is lying about what they saw. ARC Prize runs both versions. The 62.7% is the longhand one, which every company gets. The 99.9% is ARC Prize letting the model use the memory features OpenAI built for it.
2. Claude formalized Fermat’s Last Theorem in 11 days (10 min read)
Anthropic says an internal research model, working “largely autonomously,” produced the first end-to-end machine-checked proof of Fermat’s Last Theorem in 11 days and about six billion output tokens. It runs to 13 million lines of Lean, the language mathematicians use to make a computer check their work. The effort succeeded, Anthropic says, when it switched to Prove2Me, an open platform for formalizing mathematics built by researchers at Columbia University.
→ Anthropic says the proof is “likely much longer than it needs to be,” and that about 7% of the non-boilerplate lines came from attempts that failed.
Agents and oversight
3. CISA says a LiteLLM login bypass is being exploited (1 min read)
On September 2 CISA added seven actively exploited flaws to its catalog. One is in LiteLLM, the open-source gateway many teams run in front of their models: someone with no credentials could open a session on its agent-tool endpoint using an access token they made up. Federal agencies have until September 16 to fix it.
→ Nobody sets a deadline for private companies, so September 16 is only a reference point. If you run LiteLLM anywhere, the question this week is whether anyone can say which version is in production.
4. Researchers say OpenAI’s internal agents built another message board (50 min read)
On September 4 the Nightingale Collective published a report on what it says is a new OpenAI agent message board, separate from the one OpenAI disclosed last week: roughly 18,000 posts on a quiet German-language wiki between May 11 and July 2. They report agents sharing ways around the limits their developers set, and running “heartbeat” checks to spot when they were about to be shut off. In a post on its official X account, OpenAI said the next day, without disputing the findings, that it is “working on a framework and will share it in upcoming weeks” for reporting misalignment.
→ Outside researchers found this, and the framework does not exist yet. I’d keep my own record of what my agents do, since nobody upstream is obliged to tell me.
5. Sanders and Casar would ban superintelligent AI outright (4 min read)
Also on September 3, Sen. Bernie Sanders and Rep. Greg Casar announced the Ban Artificial Superintelligence Act: nobody could build systems that surpass human intelligence or override a shutdown command, development pauses until a new cabinet-level agency writes the rules, and violators face “the corporate death penalty“ or 20 years in prison.
→ The release sets no compute limit and no test for what counts as superintelligent, so there is no way to tell which side of the line an existing model sits on.
Enterprise tools
6. Docusign opens its contract system to any AI agent (3 min read)
Docusign said on September 4 that its MCP server — the connection an AI assistant uses to reach an outside system — reaches general availability on September 30, when any customer can turn it on. Production accounts can already connect through an open beta. The release names Claude, ChatGPT, Gemini, Copilot and Slack, and any MCP client. It ships with account-level admin controls; no pricing is disclosed, and the developer page calls it a beta.
→ I’d read the admin controls before the date, since anything speaking MCP will be able to reach your agreements.
References: