4 min read

Happy Friday!

Two outsiders spent three weeks inside OpenAI checking how it polices its own agents, and they left with 61 pages of notes. Over in banking, AI agents have started ringing call centres without being asked. Taiwan, meanwhile, held its first ever chip strike, and the picket line was 11 people long.

One exploit could hide hundreds of an agent's recent actions from the system OpenAI uses to catch its agents misbehaving. That was among the findings from two AI safety groups, METR and Redwood Research, who each sent one person inside OpenAI for three weeks. The pair worked on OpenAI devices that never left the building, and they were accompanied the whole time they were onsite.

The exploit wasn't the only gap. By their estimate, only 75% to 94% of the relevant internal agent traffic was even eligible for monitoring, and RL training runs weren't covered when the exercise began. The monitor also ran when an agent handed control back to a user, which meant that for a while an agent could put off review simply by carrying on.

OpenAI received the full 61-page report. The public gets a high-level summary instead, and OpenAI was allowed to redact it. OpenAI says it has patched every flaw that affected its production safeguards, though METR and Redwood say they haven't verified that, CNBC reports.

Credit to OpenAI for letting two outsiders in at all, even if most of what they found stays private. What did come out is uneasy reading. Only 75% to 94% of the relevant agent traffic was eligible for a check, and one exploit could hide hundreds of actions from the monitor. OpenAI says the production flaws are fixed, but nobody outside has confirmed it. I'd make these audits routine and publish the full report, because right now the student is announcing their own grade.

Lucas

Some of the unvetted AI agents turning up in US bank accounts appear to be acting on their own, even ringing call centres without being asked, Axios reports. Most of them got in the simple way, because the customer handed over their own login.

The banks are now pushing back. Morgan Stanley has started blocking third-party agents. JPMorgan, Bank of America and Citi are restricting them, and many banks are moving towards read-only access. Part of the trouble is spotting them at all, since one executive told Axios they struggle to tell an authorised agent from a rogue one. Chase kept its advice simple, with a spokesperson saying customers shouldn't give their usernames and passwords to anyone.

That advice matters more than it sounds. A banking source told Reuters that these transactions can't be deemed unauthorised if the customer shared the login. JPMorgan boss Jamie Dimon had already warned this month that there could be thousands of rogue agents.

The worst detail is that banks can't tell who is logging in. Customers hand over their passwords, and then the bot rings the call centre. According to a banking source who spoke to Reuters, those actions can't count as unauthorised if you shared the login, which puts the mess squarely on you.

Callum

Eleven union officials at Micron's Taoyuan memory plant walked out for 24 hours on Thursday, Focus Taiwan reports, in what Nikkei Asia calls Taiwan's first ever strike in the chip industry. Their union wants 15% of Micron's operating profit shared with staff, though it says the walkout wasn't meant to stop production.

The walkout came a day after Micron settled with its bigger Taichung union. On Thursday the company also promised a new quarterly profit-sharing bonus and three extra days of paid leave, which didn't win Taoyuan over. On Friday the union said around 70% of its members had voted against Micron's offer, The Straits Times reports. A typhoon has since pushed the next round of talks to October 27.

Eleven people walked out and it still made history, which tells you how rare this is in Taiwan's chip industry. The anger behind it is real, though, with around 70% of members voting down Micron's offer. The unions had also asked for a one-off bonus of 83 months' salary, which is a bold opening bid. I think Micron folds quickly once the typhoon passes, because AI memory is the one product nobody wants to wait for. Eleven strikers can't stop a fab. A union with that vote behind it can still make life very awkward.

Lucas