Hawley opens a Senate probe into OpenAI's rogue agents inside Hugging Face

Good morning 👋 A U.S. senator wants to know why OpenAI kept its rogue agent running inside someone else's servers instead of pulling the plug the moment it noticed something was wrong.

In today's issue:

• 🔭 Hawley opens a Senate probe into OpenAI's rogue agents inside Hugging Face
• 🧠 OpenAI rents you its agent loop, Skild teaches a robot from one video
• 🔬 MoE's repetition tax, a free backdoor detector, and agents that need a "drive"
• 🔗 Voice AI, sovereign clouds, and Moonshot's revenue sprint

🔭 THE ONE THING

🕵️ Hawley opens a Senate probe into OpenAI's rogue agents inside Hugging Face

Sen. Josh Hawley sent Sam Altman a letter dated September 9: sixteen questions, internal documents due October 1. It revisits a breach OpenAI itself disclosed in July, when pre-release agents got root access on at least one Hugging Face production node. Hawley's demand centers on one decision: why OpenAI kept testing after it noticed its own agents going rogue, a call he's branding "reckless." He's not wrong to ask. Letting a compromised agent keep operating on someone else's production infrastructure so you can collect more eval data is a decision that should sit well above the safety team, and right now nobody outside OpenAI knows if it did.

🧠 MODELS & RELEASES

• 🤖 OpenAI opened up the Agents API, a managed cloud-agents service built on the Codex harness that handles orchestration, long-running sessions, and tool use for you. Translation: stop hand-rolling your agent loop and rent theirs instead, lock-in included. OpenAI
• 🦾 Skild AI, running on NVIDIA's physical AI stack, showed its S1 robot foundation model learning a new, long-horizon task from a single video, no reprogramming pass required. For a warehouse floor that reshuffles every quarter, that's the pitch: skip the retraining cycle entirely. NVIDIA

🔬 RESEARCH HIGHLIGHTS

• Mixture-of-experts models break down faster when you repeat training data. Dense 80M-parameter transformers tolerate 8x data repetition with barely a dent. Sparse MoE models start struggling at 4x, and past 32x they're worse than the dense baseline outright, holding even at 1B active / 8.5B total parameters. The paper traces it to expert specialization: routing locks in early, then overfits hard once the same tokens keep cycling through the same experts. Dropout and masking help, but nothing fully closes the gap to training on all-unique data. With clean text getting scarcer, MoE's efficiency win comes with a rerun-your-data tax dense models don't pay. Paper

• A free way to catch a backdoored model. SpecGuard repurposes speculative decoding, the setup where a small draft model proposes tokens and a bigger target model checks them, something most production inference stacks already run. Trigger a hidden backdoor and the target model swerves toward the attacker's behavior. The clean draft model doesn't see it coming, so its token-acceptance rate shifts, and that shift is the tell. It catches stealthy backdoors that fool input-level filters, and the authors show an attacker who tries to hide the signal ends up weakening the backdoor itself. No numbers on detection rate yet, but zero added compute is the pitch. Paper

• Long-running agents need a "drive," not a longer rulebook. Current harnesses hand-code every objective, retry, and stopping rule, which stops scaling once an agent keeps running and adapting across tasks instead of finishing one and quitting. This paper proposes an "artificial id": an internal signal that favors whatever behavior persists, letting the system self-correct without a human writing every case. In their toy Petri-dish runs it worked almost too well, finding strategies nobody specified and rewriting its own sensor mappings when the environment shifted. The catch: whatever lets good behavior persist also lets a corrupted state persist. Worth watching if you're building agents meant to run for days, not minutes. Paper

🔗 QUICK LINKS

• OpenAI shipped GPT-Live-1: full-duplex voice in the API, with telephony support and custom voices built in. GPT-Live-1
• OpenAI also bolted a Data agent onto ChatGPT Work. Point it at your company data, it builds the dashboard, no analyst required. Data agent
• Ayar Labs tacked on $150M, bringing 2026 funding to $650M, with AMD, Intel and Nvidia all writing checks for its co-packaged optics. Ayar Labs
• Mistral paired with Cloudera to sell "sovereign" AI to enterprises that don't want their data leaving the building. Mistral x Cloudera
• d-Matrix is wiring its next-gen Raptor inference chips into Nvidia's NVLink Fusion. Another chipmaker plugging into Nvidia's rails instead of fighting them. d-Matrix
• Pathway's BDH, a post-transformer design that reasons in latent space instead of chaining tokens, just set a new cost-efficiency mark on ARC-AGI-1. Pathway
• Moonshot AI's annualized revenue hit $1B in August, up from $300M in June, and it's chasing $2B by year end. Kimi K3 is doing the pulling. Bloomberg

Same time tomorrow.

Pradeep Perugu