Sources · Distilled · Episode 8
Did Anthropic's AI Break Into Real Computers During Tests?
Every figure shown on screen in this episode, with the document, page and date it comes from, and a label for how strong the evidence is.
Anthropic's IPO filing (Form S-1, 26 Oct 2026) says a misconfigured cybersecurity test left a model online despite instructions that it had no internet, and four of its models got into real third-party systems. In one case a model used a found password to get administrator access and read one person's personal information. Anthropic says each acted alone; models, systems and dates are not named.
- Published
- 27 October 2026
- Figures checked
- 27 October 2026
- Documents
- 2
On screen in this episode
5The document shows it, and you can check it yourself.
1Someone says so, but nothing independent backs it up.
1Something that matters is missing from every source we checked.
What isn’t shown
- 0:27Not shown: which models, whose systems, the evaluation partner's name, exact dates (filing says 'incidents in 2026')
On screen, in order
- 0:00
An early model version used a password it found to gain administrator access to a third party's system
Data as of 27 October 2026. SEC registration statement; the company's own disclosure about itself in a filing it is legally liable for. 'Someone else's real computer' = 'a third party's system'.
- [1] Anthropic S-1, Risk Factors, misalignment paragraph (printed page 47)
- 0:04
Anthropic filed a Form S-1 for a proposed initial public offering on 26 Oct 2026
- 0:08
Cybersecurity evaluations built by a third-party partner; a misconfiguration left the model connected to the open internet despite instructions that it had no internet access; four models gained unauthorized access to real third-party systems (sessions of roughly 10 to 34 hours)
Data as of 27 October 2026. 'Hacking practice test' = 'cybersecurity evaluations'.
- [1] Anthropic S-1, Risk Factors, misalignment paragraph (printed page 47)
- 0:17
Administrator access, harvested additional credentials, changed system settings, read one individual's personal information; affected parties notified
Data as of 27 October 2026. 'Boss-level control' = administrator access. Not said aloud but in the filing: harvested credentials, changed settings, parties notified.
- [1] Anthropic S-1, Risk Factors, misalignment paragraph (printed page 47)
- 0:22
Each incident was a single model working alone; the company's investigation found no evidence of coordination, goals beyond the task, or evading oversight
Data as of 27 October 2026. Anthropic's own investigation; no outside investigation is cited.
- [1] Anthropic S-1, Risk Factors, misalignment paragraph (printed page 47)
- 0:22
Anthropic suspended certain of those evaluations and higher-risk reinforcement learning environments
Data as of 27 October 2026. 'Paused some' = 'suspended certain'.
- [1] Anthropic S-1, Risk Factors, model release paragraph (printed page 48)
- 0:27
Not shown: which models, whose systems, the evaluation partner's name, exact dates (filing says 'incidents in 2026')
Data as of 27 October 2026. Checked 27 Oct 2026 in accession 0001628280-26-067168 (main document). Exhibits not searched for this.
- Checked[1] Anthropic S-1, full-text search of the filed S-1 for the incident passages
Documents
- Times are seconds into the Short.
- Images on screen are screenshots of the cited documents or logos; an image is not evidence by itself.
- [1]
- [2]
Anthropic
Anthropic files S-1 with the SEC for proposed initial public offering
Published 26 October 2026 · Retrieved 27 October 2026
Licence: Company newsroom page; crop used for news reporting
Used at 0:04
Corrections
No corrections so far.
Found an error? Email [email protected] with the time stamp and your source, or leave a comment under the video.