Meta introduces Muse Wonder and Muse Wonder Full

Meta’s new flagships, successors to Muse Spark, lead with two agent features: Active Memory (carrying relevant details across conversations) and Active Focus (trying other approaches and recovering from errors when a task gets hard). They ship through the Meta Model API and Muse Code; there are no weights to download, and the announcement says nothing about whether any will be released. It gives no benchmark numbers, and the safety report under Meta’s Superintelligence Scaling Framework is promised “soon.” The company that let anyone download Llama’s weights now keeps its strongest models behind an API and a subscription, and ships the model before the safety report; both are worth tracking.

Paper: models can provably hide arbitrary capabilities

The construction embeds an encrypted backdoor network inside a host model; it stays dormant until the input carries a secret key, then decrypts and executes. Under cryptographic assumptions the hidden computation is hard to elicit or interpret even with full weight access, and fine-tuning, weight noising, gradient-based trigger search, and adversarial training all fail to surface it. That cuts at the core assumption of capability evals: what an eval finds is only what nobody bothered to hide, and under adversarial conditions “we didn’t find it” does not mean “it isn’t there.”

Cross-session attacks on agent skills: last run’s files are an attack surface

Files, configurations, and caches that a skill leaves behind get reused by other skills in later sessions. The paper maps three attack types (instructions planted in artifacts, data access across skill boundaries, poisoned artifacts that skew later conclusions) with attack-chain success up to 61% on Codex, Gemini, and Claude; the proposed defense, SkillRelayGuard, restores artifact provenance and adds a pre-execution advisory review, cutting success from 69.5% to 10.9% while keeping task completion at 87.5%. Prompt injection within a single conversation is the well-studied case; the authors describe what a previous session leaves on disk as largely unexplored.

X-AgentRed: red-teaming “each step compliant, the combination harmful”

A personal agent wired to multiple services can make calls that are individually fine while their accumulated effects cause harm. X-AgentRed builds a dependency graph of which tool’s effects satisfy which tool’s preconditions, then refines attack paths with execution feedback: across four frontier models it finds 1.77x the harmful paths of the STAC baseline with 2.99x the tool coverage.

Detecting subliminal learning without knowing what to look for

Subliminal learning is when a distilled student inherits a teacher’s trait through data unrelated to that trait; prior defenses required specifying the trait in advance. This paper exploits an asymmetry: the trait expresses consistently across teacher-generated datasets while task skills vary with the data, so analyzing parameter updates alone, with no labels, recovers the trait’s weight-space direction and suppresses it at almost no task-performance cost. Dropping the know-what-you-seek requirement is exactly what an audit of a distillation pipeline needs, though the paper demonstrates the method only on benchmark traits, not production pipelines.

AgentValve: payment authorization that resists manufactured reputations

In markets where agents pick providers and pay with no human review, attackers can mass-produce cheap identities and fabricate feedback (a sybil attack) to inflate a provider’s reputation. AgentValve records reputation as a trust graph with finite capacity: payments are authorized along trust paths with enough remaining capacity, reserved atomically by smart contracts, so inflating influence requires spending real money in proportion to the damage it enables. If agents are to pay each other with no human review, this is the kind of trust machinery that has to exist first.

Brad Smith: washing machines face more safety rules than AI, and the UK should lead

On a BBC podcast, Microsoft president Brad Smith said “right now, your washing machine is subject to more safety rules than AI,” noting every milk carton in a grocery store is subject to safety rules and inspection; he was also explicit that voluntary compliance fails because only the good companies comply. His proposals are concrete: incident-report obligations when an AI agent causes a breach, review by independent government evaluators before release, and emergency brakes to pause or slow systems that get out of control. He argues the UK should lead given the AI Security Institute’s technical depth and Washington’s greater receptiveness to London than to Brussels; whether any of this reaches the UK legislative agenda is the thing to watch. (Also covered by TechCrunch.)

Research radar

DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

RL fine-tuning of diffusion models has run on two tracks: reverse-process Flow-GRPO suffers poor credit assignment, forward-process DiffusionNFT is efficient but unstable without hand-tuned EMA. The paper finds a hidden forward-reverse consistency force inside NFT-style objectives and removes it by design, aggregating each prompt’s rewards into one improvement direction: comparable or better rewards at roughly one twelfth the GPU hours of either baseline, with no per-task EMA tuning. Directly useful if you train diffusion or multimodal generators with RL.

S-Space: Exploring Spatial Workspace in Multimodal Language Models

The authors locate a linear subspace in multimodal LLMs where object tokens encode continuous coordinates along horizontal, vertical, and distance axes; it is intrinsic to the model rather than induced by the prompt, and stays readable even on non-spatial questions. Causal interventions shift spatial judgments while leaving other answers intact. This turns “MLLMs can reason about space” from a black-box observation into a representation you can locate and edit.

Attention Sees, MLP Refuses: Tracing the Attention-to-MLP Safety Pathway in LLMs

The paper splits refusal into recognition and execution, tracing residual streams across 18 model pairs (Qwen, Llama, Gemma): harmfulness reads out mainly from early attention layers, while the refusal text is written mostly by MLPs, and resetting downstream MLPs to clean values erases most of the refusal shift caused by upstream attention perturbations. A combined-signal detector, Sentinel, beats single-signal baselines on both jailbreaks and over-refusals. Worth reading if you work on safety interpretability: refusal now has a concrete computational pathway instead of a works-or-fails verdict on alignment.

One line for today: the backdoor paper shows that under standard cryptographic assumptions, arbitrary capabilities can be hidden in weights where no eval will surface them. Against a developer who hides on purpose, a clean eval result stops meaning a clean model: evals can falsify, but they cannot certify.