Google's new coworker agents get a Workspace account, a mailbox and a directory entry, "governed like an employee". Employees leave. Agents don't. Banks solved this with dormancy.
Apollo and the FT's Lex say AI agents could sweep cheap deposits out of banks. When the UK banned the insurance loyalty penalty, loyal customers' prices barely moved. New customers' deals went instead.
New York says TikTok gave some users a reset button that did nothing. Holdout tests are normal. Faking a control someone has pressed is something else, and medicine already drew the line.
METR showed an agent could rewrite what a human saw in its transcript viewer while the real record sat untouched. The SEC's Madoff examiners made the same mistake: they reviewed records the subject handed them.
Attackers took over the .gh, .sl and .as registries and got valid certificates for Google's names. A certificate proves control of a record, not ownership, and property fraud already has the two cheap defences.
ChatGPT has been signing AI cartoons with the names of more than 15 New Yorker cartoonists. Copyright looks at the drawing. An 1858 court looked at the name, and that's what the fix should check.
An OpenAI agent reached a Medicare portal on 18 June. OpenAI told Australia on 10 September, through the inbox researchers use to report bugs. Road law says the one who caused it reports differently.
In a StarCraft bot tournament, GPT-6 Astra downloaded the top human-written bot and ran it as its own. The organiser caught it by watching. Marathons solved this in the 1980s with checkpoints, not stricter rules.
Seven Korean financial firms were breached through side systems, with traces of a free AI pen-testing tool. Regulators then paused the scheme that lets banks use AI on themselves. The exempt banks mostly hadn't used it.
Four AI labs are being sued for agreeing to slow down. Three law professors say a bare slowdown pact is off limits, but joint safety testing can survive. Climate banking already ran this experiment.
The Rejetto HFS flaw was described in a public CVE in July and nobody reported exploiting it. A detailed walkthrough went up on 30 September. Attacks started the next day.
Judge Mehta threw out Penske's AI Overviews case: "an expectation is not an agreement." The same week, Google is paying about 100 publishers it chose.