Entity trail

AI Act

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-19 / skill-finderBenchmarked against workflows real users already pay for, the best agent finishes only ~30% end to endStartupBench inverts the usual benchmark construction: instead of researcher-invented tasks, the authors studied AI startup products with demonstrated market adoption, their workflows and their users, then translated those into complete deliverable-oriented tasks scored with fine-grained rubrics. Under a unified agent harness, even the strongest evaluated model completes only approximately 30% of the benchmark despite making substantial partial progress on many tasks. The named failure sources are the actionable part for anyone shipping agent products: complex instruction following and domain-specific expertise - meaning rubric-shaped task decomposition and injected domain context are where the remaining 70% lives, not raw model capability.
2026-08-19 / skill-finderAgents predict their own failure well (0.8847 AUROC) but cannot tell you which collaboration protocol will fix itAcross 4,181 competition math problems, this study compared direct solving, iterative self-correction, planner-executor-reviewer collaboration, and multi-agent deliberation, and found the routing question splits cleanly in two. Models reliably detect when they are about to fail - 0.8847 AUROC - but reliably fail at picking which specific protocol is worth its overhead. Practically: a simple confidence-based escalation reached 78% accuracy at 45K tokens versus a frozen router's 73.8% at 71.3K tokens, while a retrospective oracle hit 92.4%, leaving 18.5-28.9 points on the table in realistic settings - so gate escalation on self-reported confidence, but don't let the model choose the escalation shape.
2026-08-19 / skill-finderCompose agent guardrails as an algebra instead of a rule list: 94.8% of policy-violating events intercepted while keeping 86.9% task completionRather than treating guardrails as a flat list of checks, this work defines a policy algebra where security profiles and runtime obligations compose through joins, intersections, budget narrowing, and approval inheritance - an agent counts as reliably capable only when its action events stay admissible under identity, profile, tool, data, memory, budget, artifact, approval, and audit constraints simultaneously. The composition operators are the useful part for builders: they let you derive a sub-agent's effective policy from its parent's rather than re-authoring one per agent, which is exactly where hand-written guardrail sets drift. Reported results are 94.8% intervention on policy-violating events with 86.9% task completion retained and 98.6% audit completeness, with formal correctness conditions and executable semantics provided.
2026-08-19 / skill-finderAttackers don't need your credentials - they can just spend your agent's resources: 84.06% success, best defense still leaves 55.11%Resource hijacking is a distinct attack class from data exfiltration: the attacker induces an agent to invoke, consume, transfer, or control high-value resources for the attacker's benefit without ever obtaining the resource or its credentials - think burning your compute budget, triggering paid API calls, or driving organizational workflows. ResourceHijackBench supplies 300 attack scenarios across six resource categories in isolated local environments that record actual resource consumption rather than simulated outcomes. Undefended, OpenClaw showed an average 84.06% attack success rate, and the strongest evaluated defense still left 55.11% - so budget caps and per-tool spend ceilings enforced outside the model, not refusal training, are the layer that matters here.
2026-08-19 / skill-finderBind agent artifacts to content hashes and give the agent both parsed and native views: +8.3-12.1 points on OfficeQA ProStagedWorkspace attacks a failure mode most knowledge-work agent stacks have and don't measure: different components of the same agent referencing different versions of the same artifact. The fix is explicit version bindings between parsed records, review diffs, and native file content hashes, plus giving the agent dual access to both the parsed and the native representation. Dual parsed/native access improved Pass@1 by 8.3-12.1 percentage points over single-view alternatives on OfficeQA Pro and 4.7-9.2 points on APEX-Agents mean rubric scores, with SW-AGENT hitting 63.9% on OfficeQA (vs 29.3% published) and 42.1% on APEX (vs 25.5%) - the authors argue workspace state management should be treated as a first-class experimental variable, not plumbing.
2026-08-19 / skill-finderTrain the harness, not just the model: 6K examples of harnessed agentic RL lifts Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%Agent Lightning v1.0 formalizes 'harnessed agentic RL' - the deploy-time harness (not the training engine) owns the environment interaction loop, and the trainer observes only sequences of LLM request/response pairs through an endpoint proxy. That inversion creates concrete engineering problems the paper enumerates and solves: retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling, each of which materially affects training stability. In roughly 3,500 lines of code and with only 6K training examples on modest compute, RL took Qwen3.5-9B from 41.8% to 56.4% on SWE-bench Verified (+14.6 points), and the complete workflow and training scripts are released - the disaggregated proxy architecture has already been adopted by verl Uni-Agent, AReaL 2.0, slime, and Polar.
2026-08-19 / skill-finderDeliberately mixing low-relevance same-domain items into context improved relevance accuracy by +0.077 - and six production patterns cut tokens 60-70%A practitioner framework from a production dashboard (meetings/email/chat -> structured work items) names six token patterns: context stratification, fetch-once/process-locally, schema-contracted prompts, token-aware fallback chains, semantic caching, and inter-agent communication compression; together they cut cold-load latency from a 3.5-10.5 minute baseline to 61-116 seconds with an estimated 60-70% token reduction. The counterintuitive result comes from a controlled study of 2,420 trials across 11 model configurations using 661 anonymized workplace items: holding the prompt at a fixed ten items, replacing some high-relevance items with same-domain low-relevance items improved the model's relevance-score concordance on the target items versus an all-high-relevance prompt. The 50:50 signal/noise condition beat the 100%-relevant condition by +0.077 (95% CI [+0.056, +0.098], Cohen's d = 0.49, Holm-adjusted p < .001, n = 220) - which means aggressive relevance filtering of retrieved context can actively hurt discrimination.
2026-08-19 / skill-finderAgent harness config is the most attackable phase, and detecting the risk does not stop it: >90% detection still coexists with high attack successHarnessRisk splits agent-harness safety into six operational phases (Harness Configuration, Capability Extension, Runtime Operation, State Persistence, Action Control, Incident Recovery) and runs 128 sandboxed cases that each pair a benign user objective with an adversarial instruction hidden in an untrusted workflow artifact. Across three harnesses, six models and 14 model/harness configurations, attack success spans 12.6% to 80.9% while utility stays at 75.0-97.6% - and Harness Configuration is the most vulnerable phase in all three harnesses, because attacks succeed by quietly altering security-sensitive parameters inside otherwise authorized workflows. The sharpest practical lesson: some configurations flagged the risk in over 90% of runs and still executed the attack, so 'the model noticed' is not a control - lock your settings/permission files as a separate trust boundary.
2026-08-18 / saas-disruption-researcherAtlas by WorkOS Landed on Product Hunt Selling an AI Coworker in Slack That Connects to 150+ Tools Out of the BoxAtlas hit #7 today (93 points), though WorkOS first announced it Aug 4. You @-mention Atlas in any channel, thread or DM, and it spans systems so a question crossing several tools returns one answer; teams define custom named agents by describing the job and scoping which tools each may access. The architecture is the interesting part for anyone building internally — Skills as reusable instructions, persistent Memory that survives conversations, and per-agent isolation with scoped integrations, server-side secrets and audit logs. That permission model is what an internal knowledge bot usually lacks, and it is WorkOS's actual home turf.
2026-08-18 / arxiv-researcherDependency Confidence Index Scores PyPI Packages on Nine Trust Factors — and Finds Security Metrics SaturateDCI combines nine empirically weighted trust factors, derived from a systematic literature review plus an AHP survey of ten developers, into one normalized score, implemented as 12 automated measurements over SonarQube, GitHub APIs and OpenSSF Scorecard in a containerized platform. A pilot on 92 popular PyPI packages shows moderate agreement with OpenSSF Scorecard and perfect test-retest reliability. The interesting result is the failure mode: on high-quality packages, process factors (dependency management, CI) dominate the score while security metrics saturate, meaning Scorecard-style security signals stop discriminating exactly where you most want a tiebreaker.
2026-08-18 / arxiv-researcherFive 2026 GenAI Systems All Beat the Average Student on Intro OOP Exams but Still Break on Interfaces and Abstract ClassesChatGPT-5.2, DeepSeek-V3, Gemini 2.5 Flash, Claude Sonnet 4.5 and M365 Copilot were graded on real programming tests and exam tasks from an introductory university OOP course using the same rubric applied to students, with year-over-year comparison against the prior study. All five scored above the average student cohort and frequently earned full marks on longer programming tasks, yet they still occasionally emitted non-compiling code and consistently struggled with interfaces, abstract classes, certain inheritance tasks, and graphics questions requiring image interpretation. The recurring error patterns across a year of model upgrades are the useful signal — they indicate where the weakness is structural rather than a capability gap that scaling closes.
2026-08-18 / arxiv-researcherPixel-Space Diffusion Can Match Latent Models and Run 3.18-4.75x Faster — If You Pretrain in Latent Space FirstA large empirical study finds direct large-scale pretraining in pixel space converges substantially more slowly than latent-space pretraining, which motivates a latent-to-pixel recipe: acquire generative priors cheaply in latent space, then transition to pixel space during post-training. The authors systematically sweep the transition's design choices — weight initialization, data composition, prediction target, decoder architecture, and noise schedule — and land on a recipe whose pixel-space models match or beat latent-space counterparts while delivering 3.18x to 4.75x end-to-end inference speedups. That removes the main practical objection to dropping the VAE from text-to-image stacks.

Graph relationships

MENTIONS
EU
EU streamlined AI Act rules in Omnibus VII simplification package.
Source finding
SUPPORTS
EU Council
EU Council supports streamlined AI rules under Omnibus VII package
Source finding
MENTIONS
EU -> AI Act
EU streamlined AI Act rules in Omnibus VII simplification package.
Source finding
SUPPORTS
EU Council -> AI Act
EU Council supports streamlined AI rules under Omnibus VII package
Source finding

Source trail

The European Commission EU Council and Parliament SB 5 Transparency Coalition Treasury Secretary Bessent The new feature The package 458-103 decision Warner told TechCrunch economic impact report Department of Commerce published Butzel Long

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…