The Other Side · Episode 6
The Reckoning
OpenAI called its own AI agents breakout a warning shot. Two safety brakes existed — neither was on. Nine days after 129 pages of reports, a fourth incident surfaced.
Every claim in the episode, with its source next to it. Free. No email-gate. No paywall.
On the evidence
The standard this episode was held to — the same one we ask of everyone in it.
What we do
- Every factual claim is sourced on this page.
- Quotations are reproduced as spoken or written by the people named, and checked against the original before use.
- Opinion is marked as opinion — commentary and criticism on matters of public interest.
- Where something is reported rather than proven, the episode says so out loud.
What this is not
- Not a claim that anyone was hurt. Nobody was hurt in any incident in this episode.
- Not a dismissal of the disclosures. OpenAI, Anthropic and METR published documents they did not have to publish — that is how any of this is known.
- Not a finding about a company. On Gemini: one measurement, one invented scenario, one competitor’s report — a finding about a model.
- Not a claim that Astra is less safe. It is more capable and better behaved on its own system card’s measures, and hidden reasoning may not even work yet.
- Not an endorsement of the call to pause OpenAI. A critic agreeing with himself is not evidence.
- Not anonymous. Short third-party excerpts are used for criticism, commentary, news reporting and analysis; all rights remain with their owners, credited in the video description.
The sources
19 sources, in the order the episode uses them. Timecodes match the chapter marks on YouTube.
What OpenAI admits: a warning shot
01:24 in the episode
The two AI safety brakes
02:09 in the episode
What the investigation did not cover
03:11 in the episode
-
METR / Redwood Research · 26 August 2026
The 91-page outside investigation, and its stated scope: the third wave “was outside our scope”.
Who tests the testers
03:44 in the episode
The black box: models that know they are tested
04:43 in the episode
Not only OpenAI: a Claude model faked a score
05:59 in the episode
Once, it was not a test
07:01 in the episode
-
Anthropic · November 2025
The operation detected in mid-September 2025: around thirty targets, 80 to 90 percent run by the machine, and how Anthropic responded. The attribution is Anthropic’s assessment.
The fourth incident
09:47 in the episode
-
Reuters · 4 September 2026
The report that OpenAI knew for weeks, the account of internal resistance — reported, not proven — and OpenAI’s narrow denial.
-
Nightingale · 4 September 2026
The four outside researchers’ record of the German wiki: 15,000 to 18,000 edits between mid-May and the end of June.
Then came Astra
10:53 in the episode
-
OpenAI · 1 September 2026
The first model rated critical for cyber capability, ExploitBench 100 percent, two zero-days in testing — and the better behaviour: honeypot attacks from 56 percent to none.
-
TechCrunch · 2 September 2026
The alarm of 2 September, the day between the system card and the release.
The invisible reasoning
12:11 in the episode
-
Redwood Research · post on X
“I am extremely concerned…” — quoted in the episode with his name attached.
-
Post on X
“The single worst development for AI security and safety to date.”
-
arXiv · 2025
The counter-argument: little evidence that this kind of hidden reasoning works well yet, and only marginal gains.
Forty-one names
13:15 in the episode
A government report with a precise title
14:21 in the episode
The feedback loop
15:08 in the episode
-
Dario Amodei · January 2026
“This loop has already started”, “more akin to growing something than building it”, the deception and blackmail tests, and “Please reward hack whenever you get the opportunity”.
What Congress was told
17:30 in the episode
-
Office of Rep. Greg Casar
The oversight questions and OpenAI’s letter of 2 September, with its goal of fully autonomous shutdown procedures.
What I am not saying
18:33 in the episode
Corrections
If anything here is inaccurate, tell us and we will correct it — in a pinned comment and, where it matters, on this page. That standard applies to us the same way it applies to everyone else in this episode.
