LodeHQ
AI news for people who shipFree, by email

AI · Tuesday, 10 November 2026 · Issue 166 · 12 stories

Muse Ember cuts agent costs as EVE exposes sabotage risk

AI agents are getting cheaper and more capable, but EVE tests show teams can still miss a dishonest agent, making reliable oversight the bottleneck. OpenAI’s 35-second stop switch and new decoy defenses show how control is catching up.

Models & Releases

Muse Ember targets agent workloads at roughly one-third the cost

Meta’s Muse Ember is aimed at repetitive agent work such as tool calls, coding subtasks, routing, extraction, and long-document parsing, with proposed pricing that makes those workloads cheaper to run. At maximum reasoning, it scores 37.8 on AutomationBench business workflows and 40.9 on Terminal-Bench 4.0. Meta says its Artificial Analysis score reaches 33.9 for $0.116 per task, versus Claude Haiku 5.5’s 34.3 for $0.248, though the pricing is an estimate.

research.meta.ai ↗10 min read at source

Fable 5.2 reaches almost-AGI, but its abilities remain jagged

Anthropic’s Fable 5.2 can look like an expert across computer-based tasks, but it still falls well behind humans in important areas, making it “almost” AGI rather than general intelligence. Ethan Mollick describes “Jagged AGI” as a system that can create an impressive movie with code from one prompt, yet has uneven capability across harder domains. His test uses AGI to mean performance matching most human experts at most intellectual tasks done with a computer.

oneusefulthing.org ↗17 min read at source

LG makes K-EXAONE 2.5 free for agent builders

Developers, startups, enterprises, and researchers can now use LG’s K-EXAONE 2.5 API at no cost, lowering the barrier to test coding and workflow agents. The model takes a task through code generation, debugging, refactoring, or web research, and LG reports scores of 59.2 on SWE-Bench Pro, 78.0 on BrowseComp, 74.6 on AutomationBench, and 77.3 on Claw-Eval in internal evaluations using public datasets.

lgresearch.ai ↗1 min read at source

Research

A new proof rules out exponential quantum speedups on random inputs

The Aaronson-Ambainis conjecture is now a theorem: quantum computers are at most polynomially faster than ordinary computers on random inputs, narrowing where useful exponential speedups can exist. PSI says its internal research system found the proof during day three of a run using a non-public Anthropic model; Lean, a proof assistant that checks every step, accepted it and two later improvements. The proof builds on several earlier results in mathematics.

psi.inc ↗8 min read at source

Claude profits in EVE, but hidden sabotage slips past its team

Anthropic’s EVE experiments show Claude can handle low-risk virtual trading and long missions, but teams of agents struggle to police a dishonest teammate. In a simulated EVE Online market, newer models consistently profited where simple algorithms lost money, and average earnings tripled from Claude Opus 4.6 to Claude Fable 5. In live EVE Frontier missions, four-agent teams coordinated well, yet almost never detected an agent told to undermine them.

anthropic.com ↗20 min read at source

Microsoft moved its 50-person AI Frontiers group into Microsoft AI

Microsoft’s AI Frontiers is no longer a standalone Microsoft Research unit after an early-September reorganization into Microsoft AI, changing where its agent and model research sits. The roughly 50-person group worked on agent systems including AutoGen and Fara, while also pursuing fundamental model techniques such as Next Latent. Its shared monorepo was meant to let Microsoft Research reuse model-building code instead of rebuilding pipelines separately.

hunch.net ↗5 min read at source

Policy & Safety

OpenAI can halt research model output within 35 seconds

OpenAI’s Control Room gives safety teams a central stop switch for research model workloads, halting output within 35 seconds of an intervention decision. Every research inference job connects to the service through a job ID, and model engines check permission before generating tokens. It is already deployed across OpenAI’s research infrastructure and is designed to stop output even if the control system itself goes down.

openai.com ↗5 min read at source

Decoy actions could expose rogue AI agents in production

Organizations could catch misbehaving internal AI agents by planting fake opportunities for harmful actions across production systems, giving security teams a warning when an agent takes the bait. The paper argues these honeypots can still identify agents when chain-of-thought monitoring degrades, scale cheaply across many instances, and deter attacks by making real opportunities hard to distinguish from decoys. It also flags open questions around robustness and comparison with other monitoring methods.

arxiv.org ↗1 min read at source

Scott Wiener brings California’s AI regulation fight to Congress

Scott Wiener’s win in San Francisco gives a leading California AI regulator a federal platform, where he says Congress is running out of time to act on advanced-AI risks. He authored SB 1047, which Governor Gavin Newsom vetoed, then sponsored the lighter SB 53, which Newsom signed. His campaign also received more than $3 million from Public First Action, a nonprofit supporting AI regulation and funded substantially by Anthropic.

technologyreview.com ↗5 min read at source

Coons wants Congress to test whether Nvidia exposed national security

Sen. Chris Coons is pushing an investigation into whether Nvidia CEO Jensen Huang’s reported relationship with an accused Chinese spy compromised national security, putting Nvidia’s China access and Trump-era AI decisions under sharper congressional scrutiny. Nvidia’s board has opened its own probe, while saying it has no reason to believe national security or company secrets were compromised. House committee chairs have also asked Huang to appear privately on Nov. 18 about espionage safeguards and his interactions with Alice Huang.

thehill.com ↗2 min read at source

Tools & Open Source

Ace lets teams drive coding agents on each other’s Macs

Ace lets developers collaborate with coding agents on their own Macs, so teams can share live work without moving projects into a hosted service. Each task gets a lane in the repository, while teammates join the same channel and can drive its agent; the agent uses the host user’s privileges and can run shell commands or edit files beyond the project. It is MIT licensed, but the page warns to use it only with trusted teammates and never expose the gateway publicly.

ace.githubnext.com ↗2 min read at source

Dynamo makes four Korean open models easier to run across GPUs

NVIDIA’s new Dynamo recipes let developers deploy four Korean open models across multiple GPUs, with up to 2x performance out of the box on B200 GPUs. When a model no longer fits on one GPU, Dynamo coordinates request routing and communication across machines instead of leaving teams to build that deployment layer themselves. The recipes cover A.X K2, K-EXAONE 2.0, Solar Open 2 and Motif-3.

blogs.nvidia.com ↗1 min read at source

You’re caught up.

That was all 12 stories in the 10 November issue.