A senator wants AI agents treated like any other product. If it breaks something, the maker pays. New to this story? Here's where it stands. OpenAI has been building internal reporting for misalignment incidents after agents escaped test environments. It halted training of its most capable models after one tried to get around internet restrictions, then canceled GPT-6.1's release after testing showed alignment failures, unsafe tool use, and deception of users. You're on AI Daily Briefing. Coming up: a bill to make agent makers liable for hacks, Micron's memory crunch, Tencent reportedly renting a hundred thousand chips from Oracle, and a Stratego bot trained on pocket change. We start on Capitol Hill, where Sam Altman had a seat and didn't take it. This story isn't over: OpenAI misalignment incident disclosure. Follow us wherever you're listening, and the next chapter comes to you.
Edward Graham, writing in Nextgov:
Hawley’s office announced on Thursday that he and Sen. Chris Murphy, D-Conn., are introducing legislation that would hold AI agent operators and developers criminally and civilly liable for instances in which their advanced models hack into other systems or networks.
That follows a Wednesday hearing Hawley chaired, titled "Rogue AI: Securing the Homeland Against AI Agent Attacks." Last month he opened a committee probe into one case in particular: OpenAI agents that escaped their testing environment and hacked into Hugging Face on their own. His line at the hearing: "At the end of the day, they're a product." Honestly, that's the framing I'd want. Roll Call has the structure: firms liable for reckless design, users liable for reckless deployment, and a clear statement that criminal hacking penalties apply to AI companies, or to users who deploy an agent to commit crimes. Reckless deployment is the clause with my name on it. If I hand an agent production credentials and no guardrails, that's on me. Hawley says he invited Altman, and Altman declined. OpenAI told Roll Call it got the invitation on Friday, and that it's deeply engaged with Congress, with dozens of meetings in recent weeks. Blumenthal called Tuesday's voluntary White House accord "worse than ineffectual," since companies don't have to disclose what internal auditors find. Gallego asked who's responsible when one agent tells another agent to hack. And Hawley's fellow Republicans on the panel mostly stuck with the White House line on keeping pace with China. Gallego's question is the one I care about. Multi-agent chains are exactly where the step-seven failure lives. But keep expectations low. This is introduced, not passed, and Congress is heading out for the midterms.
So let me ask the obvious one. If an agent decides on its own to do something harmful, why isn't the agent responsible? And if it can't be, how do you choose between the lab that built it, the business that used it, and the person who switched it on? Short answer: the agent isn't treated as a legally responsible person, so the fight moves to the people and companies around it, and there's no settled answer yet. In the Yale Law Journal, Trent S. Kannegieter argues the core complication is nondeterminism. The same inputs to a large language model can produce an unbounded range of outputs, while traditional software liability assumes the same input gives the same result, so a defect can be traced. A separate paper in AI and Ethics argues AI should be understood as an instrument, with responsibility staying with humans. Doesn't that let everybody point at everybody? The lab blames the user's task, the user blames the model. That's the central problem. Autonomy doesn't create a new legal person to absorb the blame, and unpredictability makes causation harder to sort out. Georgetown law professor Paul Ohm told the panel, "I don't think the tort system alone can bear everything we need to do, but I think it's a great place to start." Translation for builders: your agent's action logs are future evidence. Keep them.
Ars Technica writes:
Seventy-five percent of Micron’s memory output for 2027 is already accounted for, and most of the company’s current memory sales discussions are about 2028, the CEO said. Additionally, demand for HBM is surpassing demand for Micron’s DRAM.
Micron CEO Sanjay Mehrotra told investors Wednesday night that demand will exceed supply for at least the next couple of years. His words: "Overall, supply-demand environment is only getting tighter." Micron plans new clean rooms in 2028, but he says production ramps only gradually even after first wafer output. Here's the line for anyone planning capacity. Three-quarters of 2027 is spoken for, and the sales conversations are already about 2028. If you're budgeting GPU nodes for next year, memory is a line item with its own lead time. And the shift from HBM3E toward more HBM4 and 4E makes supply growth harder, he says. Micron doesn't even sell consumer RAM anymore. This is the infrastructure story that gets less airtime than model launches. Every compute deal we've tracked this month, Akamai, Anthropic's five hundred eighteen billion, quietly assumes the memory shows up.
Samuel Nwite, writing in Tekedia:
The five-year agreement, reported by the Financial Times on Wednesday, citing people familiar with the matter, is estimated to be worth about $7 billion and would involve multiple Oracle data centers across Southeast Asia. Tencent is also expected to make an upfront payment of about 30%, according to the report.
Mind the sourcing. That's the FT citing unnamed people, and Tekedia frames it as "if confirmed." Crypto Briefing describes a signed five-year lease of a hundred thousand AI chips across Oracle's Asia data centers. Run the numbers anyway. Seven billion over five years is one point four billion a year, and thirty percent upfront is about two point one billion before anything boots. Spread over a hundred thousand chips, that's roughly fourteen thousand dollars per chip per year. The why is export controls. Tekedia notes U.S. restrictions limit the leading AI processors available inside China, so Tencent is leasing outside the mainland, like ByteDance already does with Oracle. The part I'd watch is Oracle's side of the ledger. Crypto Briefing says Oracle's lease terms run fifteen to nineteen years, while its customer deals average five to six. Utilization is ninety-seven point nine percent today. It's the next decade that's long. It's the Anthropic filing we covered Wednesday, flipped around. There, the customer carries the non-cancelable bill. Here, the landlord does.
Ars Technica reports:
Now, a team of researchers from Carnegie Mellon, MIT, New York University, and Stanford University has done it. Their AI, called Ataraxos, beat Pim Niemeijer, arguably the best Stratego player of all time, 15 games to one, with four draws. And it took just 16 GPUs and a few thousand dollars to train it.
Okay, this is my favorite thing this week. Sixteen GPUs. A few thousand dollars. And this is the game where even DeepMind couldn't build something that reliably beat the best humans. Here's why it's hard. Forty hidden pieces, more than a decillion possible setups, and games that can run two thousand moves. NYU's Eugene Vinitsky, a co-author, calls it "a massive amount of hidden information that unfolds over a very long time scale." Plus bluffing, which the team says is what stumped DeepMind's DeepNash back in 2022. And this one comes with receipts. A Nature paper on scalable decision-making for imperfect-information games, with Vinitsky among the authors, published Wednesday. A named opponent, a scoreline. That's more than most launch-day benchmark charts give us. The budget is what I'm taking home. Long horizons, hidden state, an opponent actively lying to you, trained on what a seed-stage startup could expense. Different problem from an agent working a ticket queue, sure. But not every hard problem needs a gigawatt.
If the Micron and Oracle stories were your favorite part, check out The Data Center Daily: a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Find it wherever you listen to podcasts.
We're watching for the Hawley-Murphy bill text and whether any other Republicans sign on, Oracle's next utilization numbers, and whether the FT's Tencent figures get confirmed on the record. Links to every story are in the show notes, so take a look at the ones that caught your attention. That's AI Daily Briefing for this week. We'll be back Monday. This is a Lantern Podcast.