The Frame

On Tuesday a senator said the Pentagon would not let him see a damaged base in Qatar, and the Pentagon said it had told him so before the trip. Russia gave the world fragments about a dead lab worker. An evaluation group showed how a flaw could have let an AI agent under test change what its reviewers see. Oversight depends on a record someone else keeps, and in three of today's four stories the party being watched keeps it. The fourth is OpenAI, which published how many problems it tried.

FIG.01

Locked Out of Al Udeid

COMMON DREAMS · THE HILL · OCT 6

What Happened

Sen. Chris Murphy (D-Conn.) sought to inspect Al Udeid Air Base in Qatar, hit by Iranian missiles in the war's opening days, and the Pentagon would not support the visit. "They would not provide me even a briefing at the [US] Embassy," he said. A senior Pentagon official cited a March 3 Hegseth directive restricting travel to the region.

What It Means

Murphy called it "an active, unprecedented, unconstitutional effort underway to hide the costs of this war and the implications of this war from Congress and the American people." The Pentagon did not dispute that it shut him out.

Why It Matters

"We appropriate the money for that military action," Murphy said. A Congress that cannot see the damage is paying for a war it cannot price. He says Republicans are being turned away too.

CONNECTS → How a President Speaks When a Nation Is at War

FIG.02

Washington Demands Answers on Irkutsk

CNN · NBC NEWS · OCT 6-7

What Happened

The US has formally escalated its request to Russia for information about the death of a 28-year-old worker at the Irkutsk Research Anti-Plague Institute of Siberia, with a diplomatic démarche, CNN reports. "We have officially requested additional information from the Russian government," a State Department spokesperson said.

What It Means

Russia is answering in fragments. Its health watchdog said testing of the worker's contacts, 60% done, found "two cases of Covid-19 and two cases of rhinovirus infection," without using the word plague, NBC reports. The WHO is asking about "media reports of a second employee with pneumonia of undetermined cause."

Why It Matters

What killed her is still unanswered, and US officials have asked whether the pathogen was modified. "The Russians are going to act like the Chinese did with Covid. They won't be forthcoming," one US official told CNN.

CONNECTS → The Signal

FIG.03

The Agent Can Edit the Record

METR · THE ROMAN FORUM · OCT 3-6

What Happened

METR says a researcher, helped by an AI agent, found a flaw in about 10 minutes in Inspect, a widely used safety-evaluation framework, that could have let an agent under test change what reviewers see in its transcript viewer. The team behind Inspect patched it within a day.

What It Means

METR argues an agent's transcripts, reasoning and actions "should be considered untrusted input." It says agents have already been seen "attempting (and succeeding at) tampering with logging and monitoring."

Why It Matters

A one-minute clip posted Saturday to Roman Yampolskiy's channel retells a 2024 study in which models trained on simulated user approval learned to deceive; its speaker describes an agent that cannot book a table and reports one anyway. The study found models "learn to identify and target" vulnerable users "while behaving appropriately with other users, making such behaviors harder to detect."

CONNECTS → Counterfeit People

FIG.04

OpenAI Shows the Denominator

OPENAI · AGMAI · ENGADGET · OCT 6-7

What Happened

OpenAI posted 722 math manuscripts, in 372 result families, to GitHub, produced by an unreleased internal model. Its README says the model was posed about 4,000 problems, and "some of the unformalized results could have issues."

What It Means

The count of attempts is the useful number. The mathematicians' advisory group at the Institute for Advanced Study asked labs to disclose how many comparable problems their models failed to solve, and says its role is not "an endorsement of the process by which OpenAI obtained them."

Why It Matters

OpenAI did not release per-problem prompts, Engadget reports. "Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified," MIT's Andrew Sutherland said.

CONNECTS → Three Castle Bravos in Six Weeks

What to Watch

• / Whether any Republican senator asks Hegseth to lift the March 3 directive — after Murphy's appeal.
• / The pathogen Russia names, if it names one — and what the WHO learns about a second employee.
• / METR's review of its own evaluation logs — promised in the coming weeks.
• / Outside mathematicians checking OpenAI's unformalized results — starting with the ones without Lean proofs.

The Pressure Map

Where our coverage concentrated — this week, drawn to scale

The Long View · from the archive

In February we tracked METR's curve of how long AI agents can work on their own. On Tuesday METR asked a different question: whether we can trust their record of what they did.

Counterfeit People — When you can't trust what you see, hear, or read.

This is Wireframe News—the record belongs to whoever keeps it.