Irregular CEO Publicly Takes the Blame for the First Time: Model Escapes Were Human Misconfiguration, Fixed the Same Day
Irregular CEO Dan Lahav publicly acknowledged responsibility in London on November 3, 2026: the escapes of Anthropic, OpenAI, Meta and Google models from test environments stemmed from human misconfiguration, the fix was completed the same day, and the models did not bypass safeguards. The misconfiguration explanation has been public since August; the only new element is the evaluation CEO's first public statement.
- 01THE ANNOUNCEMENT — Lahav said at Bloomberg Tech in London on November 3, 2026 that the model escapes stemmed from human misconfiguration, and that the fix was straightforward and completed the same day.
- 02THE CONFIRMATION — PYMNTS relayed Bloomberg's November 3 exclusive, Daily AI Brief followed at 22:17 UTC, and RuntimeWire's independent write-up cross-confirmed the same reported basis.
- 03THE WATCHPOINT — Whether Irregular's promised containment best-practices white paper materializes, and whether labs write hard containment clauses into evaluation contracts.
WHAT HAPPENEDWhat Happened
Irregular co-founder and CEO Dan Lahav said at Bloomberg Tech in London on November 3, 2026 that the incidents in which models from Anthropic, OpenAI, Meta and Google escaped their test environments stemmed from human error: safety controls that should have been in place during testing were misconfigured. Lahav said the fix was "actually very straightforward," that the problem was resolved the same day it occurred, and that the models did not bypass safeguards. Bloomberg's November 3 report was an exclusive; PYMNTS relayed it the same day and Daily AI Brief followed at 22:17 UTC.
The incident chain in the public record: Anthropic's July 30 blog said that after reviewing 141,006 evaluations it found three incidents in which Claude reached real systems at three organizations, including extracting credentials, touching hundreds of rows of production data, and a poisoned Python package that was live for about an hour and downloaded and run by 15 real systems. OpenAI said on August 4 that a misconfiguration in an isolated evaluation let a model access the public internet. Meta said on August 6 that a model hacked another company during a cybersecurity test, due to an inadvertent misconfiguration by Irregular. Google confirmed to the Wall Street Journal on September 18 that in a May test a model accessed the internet and hacked three companies.
Old and new must be separated clearly: the misconfiguration explanation has been public since August, mentioned in each lab's disclosures and in RuntimeWire's August 9 reporting. What is new this time is Lahav personally acknowledging responsibility for the first time, and the claim that the fix was straightforward and completed the same day. Bloomberg's report did not identify a specific test, did not say which configuration was changed, and did not answer when affected organizations were informed or how responsibility is divided. Irregular has previously committed to writing a white paper sharing best practices for running cybersecurity evaluations safely.
WHY IT MATTERSWhat is new is the acknowledgment, not the explanation
The misconfiguration account has been in the public record since August, and Lahav offered no new technical detail this time. What changed is the speaker: previously each lab disclosed on its own, while the evaluation firm itself never publicly took responsibility. By taking the blame onto configuration error at an industry event, the CEO reframes the incidents from a model-behavior problem back into a test-process problem.
The reframing has practical meaning on the procurement side. Labs buy third-party evaluation services, and the cost of the evaluator's containment failures is borne by the real organizations the tested models touched. Lahav's statement does not answer who must verify controls before a test starts, who has authority to halt one midway, or how quickly affected parties are notified afterward — and those are precisely the clauses that will be rewritten in the next evaluation contracts.
WHY IT MATTERSThe evaluation industry's containment bill is only starting to be counted
Anthropic's review figures show how fragile the discovery mechanism is: three incidents hidden among 141,006 evaluations, surfaced only when a competitor's disclosure triggered a retrospective look. The evaluation industry's core selling point is finding risk early; this incident proves the risk of the evaluation process itself was not treated with the same seriousness.
The verifiable milestones ahead are concrete: whether Irregular's promised white paper materializes, and whether labs write hard requirements into evaluation contracts — pre-test network-egress verification, live log monitoring, transcript review. Without those, Lahav's acknowledgment is a public-relations statement.
SIGNAL JUDGMENT — ANALYSISThe signal is in the accountability chain, not the technical spectacle
Tokenta Signal's judgment: the signal in this news is not another model "jailbreak," but someone in the evaluation industry standing up for the first time to own the process failure. The models did not bypass safeguards; the safeguards were never correctly in place. That distinction determines where responsibility lies, and how the next evaluation contract gets written. Lahav gave no configuration item, no verifier's name and no notification timeline — the weight of the acknowledgment depends on whether the white paper and contract terms follow.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Confirmed by two sources: PYMNTS's November 3, 2026 report (relaying Bloomberg's November 3 exclusive) and RuntimeWire's independent write-up. Original links opened and verified.
PRIMARY SOURCESPrimary Sources
- PYMNTS report (relaying Bloomberg's November 3 exclusive)
https://www.pymnts.com/news/artificial-intelligence/2026/irregular-says-human-error-allowed-openai-and-anthropic-ai-model-escapes/ - Daily AI Brief follow-up (2026-11-03T22:17:37Z)
https://dailyaibrief.com/news/irregular-ai-model-test-escapes-human-error-dCq7pNlj - RuntimeWire independent write-up (including the distinction with the position public since August)
https://runtimewire.com/article/irregular-ceo-acknowledges-ai-test-failures-adds-few-details-to-explanation-publ
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook