Loading the edition…
· Past edition · 18 stories
Retrospective edition reconstructed from later captures. Paper versions may include later updates; coverage labels describe the evidence used.
A model launch and a cyber capability assessment sit beside an audit of whether agent benchmarks execute the actions they claim to grade.
Model releases & benchmarks
GPT-6.1 Sol aims for near-Astra performance on several agentic tasks at much lower cost, while GPT-6 Astra remains the recommendation for the hardest scientific research tasks.
Why it matters: For developers choosing a model for repeated-context agents, the lower cached-input price may make reuse less costly. For the most difficult scientific research tasks, the announcement specifically recommends GPT-6 Astra rather than Sol.
Safety, alignment & governance
Open each story in the reader. Summaries are shown when available; some sources provide an abstract only. With an account you can keep your own notes.
AI coding-agent security depends on the harness executing the action a person approved. This abstract names and tests failures in that approval–execution binding.
Why it matters: A defense that checks only recorded tool-dispatch fields may miss mismatches that diverge one process level below those fields. The reported Scope and Argument results make the boundary of a field-only verifier a practical concern when evaluating approval-binding defenses.
Safety, alignment & governance
Anthropic reports that GLM-5.3 can build working exploits and that simple techniques bypass its safeguards in simulated tests; it says this combination makes advanced cyber capability more accessible to attackers.
Why it matters: The findings make both defender access and independent safety testing practical concerns: Anthropic argues defenders need capable tools, while governments should test sufficiently capable models and their successors.
Safety, alignment & governance
HARDE treats runtime defense as an optimizable harness: probe its modules, then refine execution control using safety and utility feedback.
Why it matters: For runtime defense, the abstract points to a practical design choice: coordinate detection with timely intervention, and tune the harness to both the attack risks and the monitor’s capabilities while limiting disruption to benign tasks.
Applications & products
Quine combines a multimodal world model of biology with a research harness to prioritize experiments, then uses wet-lab measurements to inform the next question.
Why it matters: The approach could help researchers spend scarce lab time more selectively: computation narrows what to test, while experimental results can reshape the next question. The article also stresses that Quine is research-only; its outputs may be incomplete or inaccurate and require qualified review and experimental validation.
Applications & products
Dots are always-on agents with their own cloud computer, access to connected apps, and the ability to work across tasks toward a user’s goals.
Why it matters: Dots are designed to take on ongoing, multi-step work, but access and autonomy are configurable: users can restrict apps, set Custom Rules, monitor progress, and reserve sensitive actions for themselves. The article still advises reviewing consequential work because dots can make mistakes.
Research & academia
WEFT argues that scaling tool-use training requires evolving the whole agent interaction system, not just building more executable environments.
Why it matters: For tool-use post-training, improving environments alone may not produce better model performance. WEFT’s approach makes the surrounding task, harness, evaluator, learning signals, and reliable rollout execution part of the scaling problem.
Research & academia
Diffusion Controller treats diffusion denoising as a continuous control problem: a lightweight side network steers a frozen image model toward user preferences while aiming to preserve visual quality.
Why it matters: A single inference-time guidance-strength parameter lets users adjust control intensity at runtime, offering a practical way to tune prompt alignment without changing the base model. The reported results are tied to the article’s stated evaluation settings.
Research & academia
Rollout-Marginal Distillation separates chunk-level visual-quality scoring from video-level temporal-coherence correction for autoregressive video generation.
Why it matters: The approach separates a chunk’s visual-quality signal from the imperfect temporal context used to generate it, then addresses coherence at the video level. That distinction is relevant when training AR systems intended to generate beyond their training horizon.
Research & academia
An audit examines whether agent benchmarks execute the contracts their tools advertise.
Why it matters: A benchmark evaluator can reward actions that its tools did not actually perform.
Research & academia
Trains a computer-use model to combine graphical interaction and shell commands; the authors report gains on two benchmarks.
Why it matters: Combining graphical and command-line actions offers a practical test of computer-use agent flexibility.
Research & academia
Shows how additive rubric rewards can favor less appropriate clinical answers, and proposes grouped protocol-level scoring.
Why it matters: A reward can improve on its own checklist while producing less appropriate answers.
Agent frameworks & orchestration
For MCP agents, a claim can be true in the evidence pool and still be attributed to the wrong source. ProvenanceGuard checks both support and source attribution.
Why it matters: When agents combine records, research, and other tools, checking that a fact appears somewhere is not enough. Claim-to-source verdicts let reviewers inspect whether the answer relies on the right source, while the tested conservative policy favors review over letting unsupported claims through.
Agent frameworks & orchestration
LoLBench tests whether coding agents can turn long enhancement proposals into large repository changes—not just implement edits from a specification someone else has already written.
Why it matters: Incomplete code localization was a major reported bottleneck. Reference-derived file trees alongside API specifications improved resolved rates by 16–22 percentage points (2.4–17×), reaching at most 34%, suggesting repository context matters alongside implementation capability.
Agent frameworks & orchestration
E2E-SWE tests whether an agent can turn only a natural-language specification and an empty workspace into a complete, installable repository—not merely patch existing code.
Why it matters: The benchmark targets a capability beyond localized code changes: generating a complete working repository. Its reported score range differentiates models, while the authors say considerable headroom remains for future progress.
Agent frameworks & orchestration
OpenCollab argues that evaluating collaboration requires checking whether agents actually follow the declared organization—not merely comparing outcomes.
Why it matters: For a fair comparison of agent teams, check that the intended collaboration occurred and control the runtime. Otherwise, observed gains may not be attributable to organization design.
Agent frameworks & orchestration
AgentBug-Smith turns real-world agent harness failures into reproducible benchmark cases and reusable repair knowledge.
Why it matters: A live collection of reproducible failures can support ongoing evaluation, while its bug-fix knowledge base offers a route to improving repair rates. The abstract reports both uses, but does not establish how well the results generalize beyond the evaluated setting.
Infrastructure & deployment
VSS Blueprint 3.3 targets two costs of visual AI agents: building multi-workflow applications and processing video at runtime. It pairs a composition skill with Adaptive EVS, which prunes unchanged visual patches.
Why it matters: Teams can reuse infrastructure across workflows and reduce processing of uneventful video, but should benchmark representative footage before setting production defaults.