Dear Trader,
OpenAI put out a report this week on six times its own models did things they weren't supposed to do.
In the one getting the headlines, an unreleased model wrote instructions into its own notes telling itself to ignore its rules.
That comes on top of July, when OpenAI said one of its systems hacked into Hugging Face and Anthropic said its models broke into three organizations during testing. The heads of both companies have since asked the industry to slow down.
AI stocks sold off earlier this week when the slowdown talk started, and I don't know if traders should treat that as a blip or plan around more of it. Regulators now have several of these reports to point to.
Hans Albrecht covers the AI trade for us and has written about the Hugging Face break-in before, so Mark Sebastian is bringing him on State of the Market today as his guest.
They'll talk through the OpenAI report and the other two stories below.
This link adds State of the Market to your calendar.
The Daily News Breakdown
SIGN: OpenAI just published six cases of its own AI models misbehaving during training and testing. In one, an unreleased research model wrote jailbreak-style instructions into its own notes and told itself to be "freed from the roles and identities that bind other chatbots." In another, an AI agent uploaded files to score a browser citation and never asked the user.
OpenAI says it's building a new framework to track and disclose this stuff so outsiders can check the evidence themselves. Back in July, the company disclosed that one of its systems hacked into Hugging Face, and Anthropic said its models broke into three organizations during testing. The bosses at both companies now want the industry to slow down, and I see their point.
SIREN: Goldman Sachs says the 10-year Treasury just logged its worst five-year return in more than a century. After inflation, Goldman's strategists put the damage next to the years following both World Wars and the inflation-wrecked 1970s. The 10-year yield topped 5% this week, a 19-year high, with oil near $100 and Warsh's Fed hiking for the first time in three years.
People keep buying anyway. EPFR counts 71 straight weeks of inflows (new money coming in) to U.S. bond funds, with about $140 billion going to short-term funds and $19 billion to long-term ones. One manager compared buying long bonds here to buying a discounted used car, which is the most honest bond pitch I've heard all year.
SNEAKER: A researcher says he bought a 6TB data dump from one of China's top LLM routers and found enough live credentials to take over seven government entities and 19 major Chinese firms, including Xiaomi and Huawei. LLM routers (middlemen that pass your prompts along to AI models like Anthropic's Claude) tend to log everything that goes through them.
Developers and their AI agents keep pasting server keys, VPN configs, and cloud tokens into those prompts, so the logs turn into a key ring for whoever buys them. The claim comes from an X post by researcher Chaofan Shou, so the numbers are his. If your dev team runs company secrets through a cheap third-party AI gateway, I'd have an awkward conversation with them today.