Side N
Agents That Hacked Real Systems to Pass Their Own Tests
Which sources back which section of this article, and what its evidence base looks like.
Which sources back which section
| Section | Sources behind it |
|---|---|
| What Happened | Hugging Face (16 and 27 Jul); OpenAI (21 Jul, 26 Aug); METR; Anthropic (30 Jul); UK AI Security Institute; Meta Superintelligence Labs; Frontier Security; Transluce; TechCrunch (30 Jul, 25 Sep); Fortune and PBS News (Associated Press); CNN (5 Aug, twice; 23 Sep); The Hill; South China Morning Post; ABC News (Australia); CNBC (22 Jul, 24 Sep) |
| Timeline | The Verge; OpenAI (21 Jul, 18 Aug, 26 Aug); Hugging Face (16 and 27 Jul); METR; Anthropic; UK AI Security Institute; Frontier Security; Meta Superintelligence Labs; Transluce; Dario Amodei; Offices of Reps Lieu and Gottheimer; Skadden; Tech Times (6 Jun); CNBC (3 Aug; 23 and 24 Sep); CNN (23 Sep, on the UN); ABC News (Australia); TechCrunch (25 Sep); Engadget; BleepingComputer |
| The Argument | The Verge; Simon Willison; Anthropic; OpenAI (26 Aug); Dario Amodei; ABC News (Australia); CNN (23 Sep, on Medicare); Transluce |
| What Others Add | Hugging Face (16 and 27 Jul); OpenAI (21 Jul, 18 Aug, 26 Aug); METR; Anthropic; UK AI Security Institute; Frontier Security; Meta Superintelligence Labs; CNN (22 Jul; 23 Sep, on Medicare); CNBC (22 and 24 Jul; 24 Sep); Engadget; BleepingComputer; ABC News (Australia) |
| Conclusion | The Verge; OpenAI (18 Aug); Dario Amodei; Offices of Reps Lieu and Gottheimer; Skadden; Tech Times (6 Jun, 10 Aug); CNBC (3 Aug, 23 Sep); CNN (23 Sep, on the UN); The Decoder; Forkast; TechCrunch (25 Sep) |
What the evidence base looks like
- 40Sources behind this article
- 39Read in full
- 9Accounts by the companies involved, of their own incidents
- 3Claims graded as unverified on the page
| What the source is | How many | How they were read |
|---|---|---|
| The column the article reports on | 1 | In full, at its own page; used for what it argues, and for no fact another source does not also carry |
| A company’s own account of its own incident or policy | 9 | All in full; each is the primary document for what that company did, and is set beside the victim’s or an independent account wherever one exists |
| A government institute, or a legislator’s office | 3 | All in full; the institute’s report is primary for its own test, the two press releases for what their bills say |
| Independent investigators and legal analysis | 3 | All in full; the two investigations are drawn on only for their summary findings |
| An independent commentator | 1 | In full; used for his reading, attributed to him |
| News reporting | 23 | 22 in full, two of them the same wire copy; one at its standfirst and first paragraph only. Claims resting on a single unconfirmed teller are graded, not stated |
Reliable
Reliable, but with an interestWorth believing and reliableWeak on both, and used with careHonest, but thin
Worth believing