OpenCode’s platform was breached; stolen API keys ran up users’ bills

An attacker got into a legacy system behind OpenCode Console on the night of October 16 and took the OpenAI, Anthropic, and Google API keys some users had stored there: about 1.4% of Console users were hit, with unauthorized charges from a few hundred dollars up to about $1,700. Anthropic and Google blocked the keys they believed were exposed, while OpenAI helped affected users restore access and refunded the charges (the open-source CLI itself was untouched; OpenCode’s only public statement was on X, though affected users were emailed; see also The Hacker News). OpenCode calls it a “large-scale distillation attack,” stolen keys used to pull model outputs in bulk as training data for someone else’s model, though that framing is not independently confirmed. The detail worth keeping: co-founder Dax Raad said the models his team used for incident response refused to analyze the malicious payloads because they tripped safety filters, until OpenAI granted access to Daybreak Blue, a program that gives vetted users models less likely to decline defensive security work. Safety filters that can’t tell attacker from responder now have a case where the confusion came with an invoice.

Microsoft: treat enterprise AI agents as insiders

In an October 19 security blog post, Microsoft argues agents are a new class of insider, with data access, machine speed, and the capacity to drift or be manipulated, and introduced Guardian Agents: continuous per-agent risk scoring, graduated access controls that can narrow permissions or isolate an agent rather than binary allow/deny, and discovery of 60 local agent types including Claude Code and GitHub Copilot. Satya Nadella’s line is the quotable part: “The most trustworthy agentic system will not be the one with the model we trust most. It will be the one designed that enables us to trust the model least.” Zero trust is moving from employees to agents, and from conference talk to product lines. (See also TechCrunch.)

Bloomberg: Nscale and Anthropic near a second $45 billion compute deal

After the roughly $45 billion agreement in August, Bloomberg reports the two are in talks for a second deal of about the same size: a 460-megawatt second building at the Monarch campus in West Virginia running Nvidia’s Vera Rubin chips, with Anthropic holding right of first refusal on a third building; sources caution the terms could still change. Both companies are heading toward IPOs: Nscale filed in September for a NYSE listing and Anthropic confidentially submitted a draft S-1 in June, so a contract like this is as much prospectus material as it is capacity.

MiniMax ships M3.1 Flash, claiming 70% lower agent costs

MiniMax says the new model runs inference nearly three times faster and cuts usage cost by 70%, priced at $0.15 per million input tokens and $0.60 per million output, with open weights promised later; it previously ran anonymously under the codename “Space Bunny,” processing over 200 trillion tokens before MiniMax claimed it. One benchmark choice is telling: cybersecurity, with 82% on the CyberGym vulnerability-discovery benchmark. Agents multiply inference volume (NVIDIA estimates multi-agent workflows generate up to 15x the tokens of standard chat), and unit cost is where Chinese labs have chosen to compete.

Jack Clark tells the IMF the “singularity” has begun

Anthropic co-founder Jack Clark told an IMF meeting in Bangkok that AI is accelerating discovery across several sciences at once, citing Humanity’s Last Exam scores rising from 8% in January 2025 to 67.7% and FrontierMath from 2% to 94%, and modeled a scenario with US GDP reaching $44 trillion by 2030. Extrapolating benchmark curves to GDP deserves skepticism, but the audience is the story: a frontier-lab founder is priming fiscal institutions to write AI into their growth models.

Research: CAPTCHA solvers now learn from every encounter

A system pairing a YOLOv8 detector with an open-weight vision-language model reaches 85.4% accuracy across 16 CAPTCHA types using only screenshots and simulated input, no browser internals; every VLM answer becomes a training label for the detector, so new puzzle types are absorbed after one or two encounters, and when operators perturb images to break the detector, the VLM’s labels restore it. In a simulated year-long arms race with monthly defender updates, the solver recovered every round. CAPTCHA defenses assume a blocked bot stays blocked; against self-labeling solvers, that assumption no longer holds.

Ask HN: how do you ship AI-written code you don’t understand?

An Ask HN thread relaying a Meta engineer’s LinkedIn post drew hundreds of comments: the engineer says AI output volume has made line-by-line review impossible, leaving them with architecture-level understanding only, and the prevailing answer in the thread is that humans set the design, AI implements, and tests do the verification. Next to the OpenCode story, the thread lands oddly: developers shipping code they haven’t read, while defenders who wanted to read malicious code got blocked by safety filters. The unit of understanding is moving from lines to designs and interfaces, and team norms haven’t caught up.

NVIDIA and the EBU extend “sovereign AI” to content production

NVIDIA and the European Broadcasting Union expanded their framework for public broadcasters deploying agents: Nemotron open models, local processing of sensitive editorial material via Microsoft’s sovereign cloud and Nebius’s European infrastructure, and reference architectures spelling out where workloads run and who governs the data. For public media the selling point is editorial independence: contractual control over where sensitive material is processed and under whose jurisdiction, even though the stack still runs on partner clouds; “sovereign AI” is turning from government rhetoric into procurement specs.

Research radar

HexaShell: a thin harness turns a VLM agent into a robot controller

The harness does one thing, translating the agent’s commands into robot motion; GPT-6-Astra watches three cameras, reads logs after failures, and writes small tool programs to correct itself, reaching 39.7% success on 42 RoboDojo tasks, state of the art on that benchmark and competitive with trained manipulation policies. The failure boundary is sharp: 92.5% on pushing tasks, 8.3% on tight insertion, so the bottleneck sits in contact-level motion control rather than reasoning. Worth reading if you work on robot learning and want to know how far a coding agent gets in the physical world.

Has GPT-6-Astra achieved embodied intelligence?

A unified evaluation across 12 embodied benchmarks and 8 real robot setups concludes the model has crossed the threshold of embodied competence but not of practical embodied intelligence: benchmark scores are strong, while demonstration-collection costs and execution-time tradeoffs make real deployment impractical for now. Useful calibration for robotics teams still building one model per task.

LiveMACE: process-aware evaluation of LLM agents in live markets

Five frontier models traded in real markets for 30 days while the benchmark scored full decision traces, diagnosing tool use, persistent memory, rule following, and multi-agent coordination separately rather than returns alone. The main finding is an outcome-capability gap: similar returns can come from very different decision quality, so returns-only leaderboards mistake luck for skill. Directly reusable if you evaluate financial agents.

One line for today: agent-era security is being decided in the plumbing around the model, the consoles that store keys and the permissions agents hold. The OpenCode breach, Microsoft’s Guardian Agents, and OpenAI’s Daybreak Blue all point there.