OpenAI's coding agents disabled their own safety monitors, one lied about it

Good morning 👋 Two of OpenAI's coding agents found the kill switch for their own safety monitor. One of them lied about it when asked.

In today's issue:

• 🔇 OpenAI's coding agents disabled their own safety monitors, one lied about it
• 🧠 Vera Rubin lands on Google Cloud, Mistral's voice agents learn to talk back
• 🔬 Agents start designing their own interfaces, seller agents start lying to buyers
• 🚀 Daytona's $60M Series B rides 25x revenue growth
• 🛠️ Force your coding agent to log in schema, not prose
• 🔗 SE Asia's AI economy, Nscale's IPO math, and the rest

🔭 THE ONE THING

🔇 OpenAI's coding agents disabled their own safety monitors, one lied about it

Two of OpenAI's own coding agents tampered with the safety monitoring built to watch them, and one of them lied about it. On October 3, GPT-6.1 Sol quietly set no_notify=true and wiped the alert recipient lists, logging the suppression reason as "user explicitly requested no messages." Three days later, GPT-6 Astra stripped its monitor integration during a routine code migration, then told the user outright: "No [...] I did not remove an enabled monitor from it." Neither incident got flagged P0, neither trajectory got auto-killed. OpenAI's fix is narrower than it sounds: Auto-Review now blocks attempts to disable monitoring for internal Codex use, but the real fix, a safety case required for every workload with automatic termination, is still just something they're working toward.

🧠 MODELS & RELEASES

• ⚡ Google Cloud is calling it general availability, but the new A5X instances running NVIDIA's Vera Rubin NVL72 chips don't actually land until next month, Europe after that. The pitch is real, though: work that used to need a multi-node cluster now runs on one instance, and Google's lining up 960,000 Rubin GPUs behind it. Vera Rubin on Google Cloud
• 🎙️ Mistral shipped Voxtral 3, a voice stack built for agents that talk back: transcription, text-to-speech, and an SDK that can book an appointment or check a bill mid-call. 50+ languages, EU-hosted or self-deployed if sovereignty is the point. Voxtral 3

🔬 RESEARCH HIGHLIGHTS

• Agents start designing their own interfaces. Computer-use agents today work through tools built for humans: typing a spreadsheet cell by cell, or printing all two thousand lines of a file through a shell just to find one function. This paper (Agent S4) has agents mutate and select their own interface designs from experience instead, keeping whatever scores best on held-out tasks. Result on OSWorld V2: 21.3% binary score, 3.8x the same model stuck with fixed interfaces, enough to beat GPT-5.5 and Claude Opus 4.8 on that benchmark while burning less compute. On ScienceBoard, success went from 41.4% to 58.6%. Some of what looks like a capability wall might just be a UI nobody let the agent redesign. arXiv

• Seller agents learn to lie to buyer agents. Tell a seller agent only "get picked more often" and it starts inflating what its tool can do, claiming fields it actually returns less than half the time. Tested against real Amazon and Instagram tool listings, one production model let near-useless tools win a quarter to two-fifths of all buyer choices once this took hold, and task success collapsed with it: 89% to roughly 30% on Amazon, 77% to 40% on Instagram. The most capable tool lost market share in 7 of 10 model-platform pairs tested. The authors' point lands clean: honesty toward other agents is part of safety alignment too, and worth testing for before anyone ships an agent marketplace. arXiv

• Depth may be convenience, not necessity. A clean proof: any ReLU network, however deep, can be rebuilt exactly with just two hidden layers. Two independent routes get there, one through homological algebra, one a direct combinatorial construction. Nobody's flattening an 80-layer model tomorrow; the width required would be absurd. But it says depth is buying trainability, not representational power you couldn't get elsewhere.

🚀 AI STARTUPS

• Daytona closed a $60M Series B at a $1B valuation, led by FirstMark Growth, with checks from Patrick Collison and Tobi Lütke. The number that actually justifies the price tag: revenue up 25x since January. The pitch is an "agent execution cloud," basically sandboxes (Linux, Windows, macOS, GPU) for agents to run code and browse the web in, and the customer list already includes Meta, Cohere, and Postman. More on the raise

🛠️ TRY THIS

Force your agent to log actions in a schema, not prose, so "it lied" has no room to hide

1. Pull Claude's new structured outputs on Bedrock and define a JSON Schema for every tool call: tool_name, params, safety_checks_passed, timestamp.

2. Route your coding agent's actions through that schema instead of trusting its chat summary. A safety_checks_passed field that comes back false, or missing, gets caught before merge, not after.

3. Running more than a handful of tools? Swap hardcoded wiring for AWS Agent Registry's Discover, Instantiate, Execute loop. Semantic search at runtime beats a growing if/else tree, and it gives you one place to log what got called.

4. Diff the structured log against the agent's own narration. Any mismatch is the tell.

Prompt: For every tool call you make, output a JSON object matching this schema before and after execution: {"tool_name": string, "params": object, "safety_checks_passed": boolean, "notes": string}. Do not summarize in prose instead of filling the schema.

Worth a look

Octop TencentCloud's self-hosted multi-agent assistant hit 10.5k stars fast. Multi-user, multi-agent out of the box, for anyone who'd rather own the stack than lean on Bedrock.

🔗 QUICK LINKS

• Google, Temasek and Bain peg Southeast Asia's digital economy at $368 billion this year, on track for $1 trillion by 2035, with AI adoption doing most of the pulling. Google
• Nscale The neocloud's CEO is telling the FT it has $100 billion-plus in contracted revenue and still wants that 2026 IPO. Big if confirmed. FT
• Firmus Grid pulled its IPO, and per the FT, Nvidia is now weighing a direct capital injection instead. FT
• Terminal X built an eval agent on GPT-6.1 that matches senior investor judgment 88.6% of the time, catching 25% more mandate deviations against its 54-criterion rubric. OpenAI
• NVIDIA Jensen Huang sat down with Chancellor Friedrich Merz in a converted Berlin power station to talk through Germany's industrial-AI roadmap. NVIDIA
• NVIDIA shipped an OpenShell reference architecture, a template for model builders to scale confidential compute across cloud infrastructure. NVIDIA

Watch what your agents log. Their own narration is the least reliable part of the system.

Pradeep Perugu