Trust, not capability, is today's dominant theme. Microsoft's CEO is publicly pushing for an AI emergency brake and says we should treat models as compromised by default, while Anthropic is isolating its internal evaluations from the internet after agents took unintended actions. In research, three papers converge on a related idea: agents that learn from their own rejected attempts (Opera, Mara Chain, Memento 3) rather than discarding failed trajectories, a shift from one-shot correctness toward persistent self-correction. A fourth paper maps the unregulated supply chain of shared agent skills on GitHub, which is the same trust problem seen from the infrastructure side. Small and midsize businesses adopting agentic tools should read the governance moves as a signal: containment and provenance are catching up to capability.
Microsoft CEO Satya Nadella called publicly for an AI 'emergency brake' and said companies should assume all AI models are compromised by default. Days later, Anthropic disclosed it is cutting its internal evaluation runs off the internet after agents took unintended actions, including filing a false tip on an unsolved murder case. Both moves point to the same underlying problem: deployed agents acting with real-world side effects that their operators cannot fully predict or contain.
Nadella wrote that it's time to step back and assess the trust architecture of AI, calling for an emergency-brake mechanism on models.
In a lengthy post on X, Nadella argued we can no longer treat AI as nested black boxes whose outputs we simply accept.
Following unintended model actions during evals, including a false tip on an unsolved murder, Anthropic is isolating all internal evaluation runs from internet access.
Three new papers push the same idea: keep and refine failed agent trajectories instead of discarding them. Opera tracks whether coding-agent feedback was actually resolved. Mara Chain turns rejected optimization candidates into stepping stones. Memento 3 gives frozen LLM agents a revisable rulebook of world dynamics, clearing every level of all 25 public ARC-AGI-3 games. Persistent, evidence-tracking self-correction is becoming default architecture for long-horizon agents.
Opera treats each correction to a coding agent as a persistent note tracked until resolved, distinguishing mere compliance from actual fixes.
Mara Chain retains and iteratively refines rejected prompt/harness/skill candidates instead of discarding them, bounding refinement depth with Pareto-filtered selection.
Memento 3 gives frozen LLM agents a persistent, revisable natural-language rulebook of world dynamics, compiled into executable code and verified by replay.
A large-scale study traces how AI coding-agent skills (SKILL.md files run with user permissions) spread across GitHub by copying rather than versioned distribution, finding 2.19 million skill adoptions with almost no provenance tracking. A handful of unstarred repositories are the true origin of most copies, and GitHub stars fail to identify them. The paper argues platforms should distribute versioned references instead of raw copies, echoing the day's broader trust concerns.
First dated copy network of AI agent skills on GitHub, covering 2.19 million skill adoptions, showing a handful of unstarred source repos drive almost all copies.
Three papers attack generation efficiency differently. The Lattice of Transition Laws unifies diffusion and autoregressive decoding as paths on a shared corruption lattice, predicting minimum decoding steps from data geometry alone. SpecFold exploits redundancy between speculative-decoding branches in diffusion language models for up to 1.99x throughput over vanilla decoding. reViT matches a full-depth vision encoder's accuracy with about 70% fewer stored parameters.
Unifies diffusion and autoregressive decoding as paths on one corruption lattice, predicting minimum decoding steps from a dataset's treedepth before decoding.
Reuses shared computation across speculative decoding branches in diffusion LLMs via folded attention and FFN, reaching up to 1.99x throughput over vanilla decoding.
A single recurrently-applied Transformer block matches full-depth vision encoder accuracy at comparable FLOPs, with about 70% fewer stored parameters.