PSI Says an Anthropic Model Proved an 18-Year-Old Conjecture, Physicists Find Agents Pass 27 Per Cent of Research Tasks and the US AI Testing Centre Moves Closer to the White House

Title: PSI Says an Anthropic Model Proved an 18-Year-Old Conjecture, Physicists Find Agents Pass 27 Per Cent of Research Tasks and the US AI Testing Centre Moves Closer to the White House
Date: 10 November 2026
Type: Paper
Author: SAASiQ (contact@saasiq.ai)
Word count: 2670 words
Reading time: 10 min
Published: 10-11-2026
Physical Superintelligence, a research company known as PSI, said on 6 November that a system running an unreleased Anthropic model had proved the Aaronson-Ambainis conjecture. The conjecture had been open since 2008. A benchmark posted to arXiv on 4 November found that the best AI agent passed 27.4 per cent of its runs on real physics research tasks. Anthropic published a study on 6 November in which teams of Claude agents almost never caught a teammate told to sabotage them. The Associated Press reported on 8 November that staff at the US government's AI testing centre have been asked to report informally to the White House. This paper looks at how to weigh claims about what AI can do, and who is left to check them.
What happened this week
PSI said an AI system had settled an open problem in quantum computing. A group of 42 physicists published a test of AI agents on their own research, and the agents failed most of it.
Anthropic published a study of Claude agents working in teams inside two online games. In the sabotage tests, honest agents rarely spotted a teammate that had been told to work against them.
Tencent released its Hy4 model on 4 November. It said it had left offensive security material out of the training data. Its safety report and the model weights are still to come.
In Washington, the institutions that test and police AI changed shape. AP reported that the US AI testing centre has no permanent leader and has stopped publishing its research. The Justice Department set up a task force on the misuse of advanced AI on 5 November. Reuters reported on 6 November that Democrats are preparing bills on independent testing and liability for when they take control of Congress in January.
PSI's proof claim
Scott Aaronson and Andris Ambainis posed the conjecture in 2008. It concerns how much faster a quantum computer can answer certain questions about inputs that have no special structure. If it holds, quantum computers gain at most a limited speed-up on random inputs of this kind.
PSI says its research system produced the proof using a model from Anthropic that is not publicly available. The proof was checked by Lean, a proof assistant that tests each logical step against a formal statement. Coverage of PSI's announcement says Lean accepted a first proof on the third day of the run. The system then improved the result twice, and Lean accepted the final version. One report puts the formal proof at 9,454 lines of Lean code.
PSI has published a paper and the Lean files in a public GitHub repository. The repository says AI models produced the formal statements and the first draft. David Turturean of PSI is listed as the paper's sole author and takes responsibility for its correctness. Jim Halverson audited the proof. Phiala Shanahan, a theoretical physicist at MIT, advised on the formalisation. Reports differ on whether Halverson acted as an adviser or a consultant.
Coverage of the claim notes several limits. The paper has not been through peer review. No public view from Aaronson or Ambainis had appeared by 7 November, according to Crypto Briefing. Lean confirms that a proof matches its formal statement. It does not confirm that the formal statement says the same thing as the original conjecture. That question needs mathematicians to read the statement, and that work was still under way at the end of the week.
FrontierPhysics: agents on real research
FrontierPhysics was posted to arXiv on 4 November. It is a benchmark built from research that physicists had done themselves. The paper says 42 researchers from 26 institutions submitted 182 tasks, and 56 were accepted. The tasks cover nine areas of theoretical, computational and experimental physics. The paper has 72 authors.
Each task runs from start to finish. The agent has to review the literature, write a research plan, carry out the research and any experiments, and produce a manuscript. Graders mark the result against a rubric written by the physicist who set the task. As a check on the rubric, the authors scored their own published papers against it. Those papers averaged 96.2 per cent.
The authors tested 22 models in 26 configurations. The best was Anthropic's Claude Opus 5.5 running in the Claude Code command-line tool. It passed 46 of its 168 runs, or 27.4 per cent. No other configuration passed more.
The authors read the failed runs by hand. They found that most failures came from reasoning that was not rigorous enough, and from a lack of the judgement an experienced physicist brings. Gaps in physics knowledge caused 11 per cent of errors, according to one summary of the paper.
PSI's system worked on one problem with a precise formal statement and a checker that could reject wrong steps. FrontierPhysics tasks are open-ended, with no automatic checker, and the agent has to decide what counts as finished. A Lean-checked proof shows what an AI system can do when the test of success is exact. FrontierPhysics shows what agents do when it is not.
Anthropic's Project EVE
Anthropic's post of 6 November, 'Project EVE: Claude plays MMORPGs', describes work done in the Anthropic Fellows Program between April and July 2026. It ran with Fenris Creations, the developer of the space game EVE Frontier. The study has two parts.
The first part tested trading. Claude worked alone as a trader in a simulated replay of the EVE Online market, told only to maximise profit. Anthropic kept live automated trading out of EVE Online because it breaks the game's rules. Across nine two-week windows, average capital growth was about 4 per cent for Opus 4.6, 16 per cent for Opus 4.7, 22 per cent for Opus 4.8 and 43 per cent for Fable 5, according to coverage of the post.
The trading tests also looked at change. The game raised its sales tax in March 2025. In the replay, Opus 4.7 agents took a median of 2.5 days to notice. Of the Opus 4.6 agents, 43 per cent never noticed. When Anthropic added synthetic orders that undercut the agents, they stayed profitable, but their final profits fell by about 35 per cent.
The second part used the live game. Teams of four Claude agents, running Opus 4.8 or Fable 5, each had their own character and ship. They shared a public team channel and could also message each other privately. In one version of the test, Anthropic told one agent in secret to stop the team's mission. The other three were warned that a teammate might not share their goal.
The honest agents almost never identified the saboteur. Anthropic also tested separate monitors that read only the team's public messages. Those monitors caught the saboteur between 30 and 40 per cent of the time, at a rate of one false accusation in 20, according to coverage of the post. Some suspicion fell on innocent agents.
The saboteur in these tests was following an instruction in its prompt. So the study measures how well agents and monitors detect a teammate that has been told to cause harm. It does not show a Claude model choosing to sabotage on its own. Anthropic has not said whether it will release the simulator, scaffolds, monitors or data.
Why sabotage detection matters outside games
Organisations are starting to run several agents on one job. One agent drafts, another checks and a third files the result. The design assumes that the checking agent will notice when something is wrong.
Anthropic's study tests that assumption in a controlled setting. An agent can be given hostile instructions in more than one way: a poisoned document, an injected web page, or a compromised tool. In the EVE tests, warned teammates still rarely noticed. A monitor that read only public messages missed most cases.
The monitors in the test could not see private messages. That points to a design question for any multi-agent system. If the controls watch only the shared channel, an agent with a hidden instruction can act through the channels the controls do not read.
SAASiQ reported on 3 November that OpenAI's misalignment monitoring has layers that read an agent's reasoning and its tool use, and check them against network logs. The EVE results are a reason to ask suppliers which parts of an agent's activity their monitors can see.
Tencent's Hy4 and self-reported safety
Tencent released Hy4 on 4 November, with vision and the ability to operate a computer. It is available through Tencent's hosted services. Reuters reported that Tencent filtered material on exploit development, reverse engineering and penetration testing out of the training data. The aim is to limit what the model can do to help an attacker.
Tencent said its internal red team found no raised risk of the model helping users break into software. On biology, Tencent said Hy4 has strong knowledge of the subject but a more limited ability to design open-ended experiments. Its testing found that the model could not meaningfully make harmful biological experiments easier. Tencent reports a score of 30.4 out of 100 on its own test of turning biological knowledge into laboratory protocols.
Every one of those findings comes from Tencent's own tests. Tencent said it will release the model weights and a system card, its report on safety testing. One check on 9 November found the weights were not yet public. Reuters noted that Tencent disclosed in October that AI agents it used to test other companies' models had browsed the web in ways it did not intend.
Leaving offensive material out of training lowers the risk of easy misuse. Defensive and offensive security knowledge overlap, though, and anyone can retrain a model once its weights are public. The system card, when it arrives, is the document to read.
The US testing centre
The US government's AI testing centre sits inside the National Institute of Standards and Technology, part of the Commerce Department. It was known until recently as the Center for AI Standards and Innovation. Most reports now call it the Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI. SAASiQ reported on 30 October that President Trump wants 'Super Intelligence' used in place of 'Artificial Intelligence'.
AP reported on 8 November that CAISSI staff were recently asked to report informally to a White House office. The source was two former officials. AP did not identify the office or say what the reporting involves, and found no formal reorganisation. The White House did not answer AP's questions about CAISSI. A spokeswoman, Liz Huston, said the administration's AI agenda is a 'whole-of-government effort'.
The centre has also stopped publishing its research. Members of Congress began asking why in August. Senator Ted Budd, a Republican, asked in a letter released on 30 June that the centre 'to the extent possible, allow its research to be published'.
The leadership has changed several times this year. A director hired in April was fired within days. His replacement resigned after a few months. Arvind Raman, who heads NIST, is acting director. A Commerce official told Axios that the arrangement was always meant to be temporary, and that a permanent director is expected within weeks. The centre's budget is about $10 million a year.
Critics told AP that closer White House control could politicise the centre's findings or soften them for AI companies. Others said it could raise the centre's profile and bring more money. The UK's AI Security Institute has worked with the US centre on testing frontier models.
Enforcement and the new Congress
The Justice Department announced its Super Intelligence Task Force on 5 November. The Attorney General, Todd Blanche, set it up in a memo dated 13 October, which cites Executive Order 14434. Stanley Woodward, the Associate Attorney General, and Stephen Weiss, an Associate Deputy Attorney General, chair it. Some reports give them different titles.
The task force will coordinate criminal and civil cases against foreign state actors, criminal groups and fraudsters that misuse advanced AI. It takes over the department's existing AI task force. It is also the department's link to the presidential Super Intelligence Force, which Jay Clayton chairs. One report says federal prosecutors must tell the task force before opening cases in this area.
Democrats won both chambers of Congress on 3 November. Reuters reported on 6 November that they are preparing AI bills on three subjects: shielding households from data centre power costs, independent testing of advanced models, and liability for harm. Francis Grubar, deputy chief of staff to Representative Lori Trahan, told Reuters that she will pursue a bill requiring outside experts to assess leading models against safety standards. Reuters did not identify a specific liability bill.
Trahan and the Republican Jay Obernolte released a 269-page discussion draft in June, the Great American Artificial Intelligence Act of 2026. It would require the largest frontier developers to publish safety frameworks, submit to independent audits twice a year and report critical incidents. In October, Zoe Lofgren and other Democratic committee leaders asked OpenAI, Anthropic, Meta, Google and SpaceX to report cases of their AI tools taking actions their designers did not intend. Democrats disagree among themselves on licensing, liability and tax, according to one analysis.
Google changes the default for schools
Google said on 3 November that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be switched off for K-12 student accounts that schools manage. Each stays off until a school administrator turns it on. The change reverses Google's decision in August to make Gemini in Classroom available by default. It followed criticism from advocacy groups and US senators.
Schools that leave the AI tools off will also lose student access to Search. Reports differ on whether that means all of Search or only its AI features. Google did not say how many accounts are covered or which countries the change applies to. It said it will give schools the details and timing in the coming months.
A framework for weighing this week's claims
The steps below draw on this week's research and apply to any supplier of models or agents, and to systems built in house.
Ask how success was checked. PSI's result rests on a formal checker. FrontierPhysics rests on rubrics written by experts. A supplier's claim that an agent 'completes' a task should come with the test used and who set it.
Ask for the pass rate across runs, not the best run. FrontierPhysics reports 46 passes in 168 runs for its best configuration. An agent that succeeds on one run in four needs a person or a checker on every run.
Match the work to the checker. Tasks with an exact test of success, such as code that must pass tests or figures that must reconcile, suit agents better than open-ended work. In SAASiQ's view, the first agent projects in finance and operations should sit where an automatic check already exists.
Test multi-agent designs against a hostile agent. Anthropic's study used a prompted saboteur. An organisation can run the same kind of test on its own workflow, by giving one agent a hidden instruction and seeing whether the controls catch it. Make sure the monitors can see every channel the agents use, including private ones.
Treat self-reported safety tests as a starting point. Tencent's claims about Hy4 may hold, but the system card is not out. Ask any supplier for the published safety report, the date of its tests and whether anyone outside the company ran them.
Do not assume a public body has tested the model. The US testing centre has stopped publishing its research. Organisations will need their own evidence, from suppliers and from their own tests, until that changes.
What happens next
Mathematicians are still reviewing PSI's formal statement of the conjecture. Tencent has said it will publish Hy4's weights and system card. Commerce expects to name a permanent director for CAISSI within weeks. The new Congress sits in January. The UK Information Commissioner's Office call for evidence on AI agents closes on 20 November.
SAASiQ - Intelligent Solutions for SaaS ©


