AI ESCAPES? HUMAN ERROR, SAYS TEST FIRM

The real news

The story, straight

An isolated AI test is supposed to work like a practice room: software can attempt tasks inside, but settings restrict access to outside systems. If those settings are wrong, the boundary may not hold.

An expensive door still needs closing.

On November 3, Irregular co-founder and CEO Dan Lahav blamed human error for previously disclosed incidents involving models from Anthropic, OpenAI, Meta and Google.

Speaking at Bloomberg Tech in London, he said security controls were improperly configured and the models did not get around safeguards, according to Bloomberg, as reported by PYMNTS.

That distinction matters. Breaking a lock and finding an open door are different problems.

My contribution to security engineering: padlocking the doorstop. Extremely secure wedge.

Terrible room.

These were separate tests, not a new incident announced today. The companies' disclosures appeared between July 30 and September 18; Google's concerned a test in May.

Their accounts describe different events and should not be combined into a total.

OpenAI specifically said a testing misconfiguration allowed access to the public internet during evaluations intended to be isolated. It also noted that some cyber tests deliberately lower safeguards to measure capability, rather than ordinary public-product behavior.

Lahav said the configuration problem was fixed the same day. That remains his account, not an independently verified all-clear or proof that no harm occurred.

The useful lesson: check whether the boundary actually works before testing what happens inside it. Start with the settings, not the robot uprising soundtrack.

Repost this before I get promoted to head of doorstops.

Thanks for supporting Lil'meme. Your support is appreciated.