Smart search ranks by meaning as well as keywords (one row per story, last 45 days).
OpenAI is winding down its contract supplying models to Cursor after Cursor's acquisition by SpaceX.
OpenAI announced it will wind down the contract that provides OpenAI models to the Cursor coding tool. The decision follows Cursor's acquisition by SpaceX. Developers relying on OpenAI models inside Cursor will need alternatives as the contract ends.
TrED formalizes transfer learning for domains whose data availability evolves over time, arguing classical settings are regimes along one trajectory, and remains unsolved.
The paper introduces Transfer Learning for Evolving Domains (TrED), formalizing transfer learning as a trajectory problem where target data and labels are progressively collected. TrED is specified by a data availability process fixed by the environment, a freely chosen learning protocol, and an evaluation criterion scoring the whole trajectory of models. Classical settings like domain generalization, domain adaptation, and multi-domain learning are recovered as regimes within this framework. The authors survey the literature and find most methods are tailored to a single regime, leaving TrED a well-posed open problem.
Fortinet acquired Virtue AI to add agent red-teaming, MCP scanning and runtime guardrails to its AI security portfolio.
Fortinet acquired Virtue AI to extend its Security for AI strategy beyond FortiAIGate, which protects LLMs from prompt injection, data leakage and model poisoning. Virtue AI brings automated agentic red-teaming across 50+ sandboxed environments and 14 domains, agent discovery and governance including MCP tool scanning, continuous AI validation, and real-time guardrails across text, images, video, audio and code. Financial terms were not disclosed and the consideration is immaterial to Fortinet; Gartner projects the AI security market to grow from $2.8B in 2026 to $16.4B by 2030.
A distillation framework compresses LLM reasoning into a 15.5M-parameter trade-up recommendation model reaching AUC 0.941 with product-type test-time training.
The paper targets trade-up recommendation, which identifies higher-quality alternatives that preserve customer purchase intent. A retrieval-augmented few-shot LLM teacher generates labels and rationales that supervise a compact embedding-pair classifier; at inference the 15.5M-parameter student uses only two precomputed 768-dimensional embeddings with no LLM calls. On 8,352 annotated pairs, label-only training scored AUC 0.912, reasoning distillation reached 0.924, and product-type test-time training lifted it to 0.941 with average precision 0.940. The distilled student is roughly 5,000x faster and 10,000x cheaper than direct LLM inference on a 100K-pair proxy catalog.
A 2026 scorecard ranks DSPM tools with Wiz and Cyera tied first, documenting consolidation via Palo Alto, Rubrik, Proofpoint, and CrowdStrike acquisitions.
The article ranks ten DSPM platforms: Wiz and Cyera tie at 8.7/10, followed by BigID at 8.5 and Securiti at 8.4, scored on discovery breadth, classification accuracy, access context, remediation, and value. It highlights heavy market consolidation, noting Dig Security was acquired by Palo Alto Networks, Laminar by Rubrik, Normalyze by Proofpoint, and Flow Security by CrowdStrike. Buyers are advised to purchase from current owners and confirm post-acquisition integration state.
Researchers model curriculum learning as Wasserstein transport over difficulty distributions, finding curriculum benefits are strongly task- and budget-dependent with no dominant strategy.
The framework represents curricula as trajectories of training distributions over discrete difficulty levels, decoupling ordering, matched exposure, endpoint smoothness, and pacing. Across a calibrated suite of 12 tasks and 33 difficulty axes under fixed training budgets, no single strategy dominates, though easy-to-hard ordering improves hard-level performance relative to exposure-matched static sampling. Endpoint smoothness and pacing substantially affect where along the difficulty spectrum a curriculum is effective, and the transport view supports extensions to learned pacing and structured difficulty spaces.
Travis Kalanick's Atoms, which raised $1.7B led by Andreessen Horowitz with $100M from Uber, is reportedly exploring robotaxi technology and acquisitions.
Atoms raised a $1.7 billion round led by Andreessen Horowitz earlier in the summer, and Uber has invested $100 million. The Financial Times reports the startup is preparing a hiring spree and acquisitions to become a major autonomous vehicle player, and has discussed Uber using its robotaxi technology. Atoms previously acquired Pronto, an autonomous mining startup led by Anthony Levandowski, Uber's former self-driving chief.
Researchers release OR-Clarify, a benchmark testing whether LLM agents ask clarifying questions before formulating optimization models from incomplete requests.
OR-Clarify evaluates pre-formulation clarification in operations research: each task gives a partial problem description, withholds structured hidden slots, and scores agents via bounded interaction with a simulated user, measuring slot recovery, stopping behavior, silent assumptions, and interaction cost. The authors also propose InterOPT, a two-stage framework that identifies formulation-critical gaps to decide when to ask or stop. In choice-based experiments InterOPT substantially outperforms all baselines in exact slot recovery and remains competitive in the open-ended setting.
An editorial scorecard ranks the top ten Google Cloud security tools for 2026, with Wiz, Sysdig, and Security Command Center leading on correlation, GKE runtime, and native depth.
The roundup evaluates ten GCP security tools across five weighted criteria including GCP-native depth, correlation, runtime protection, multicloud parity, and value. Google Security Command Center is positioned as the included native floor, while Wiz and Sysdig top the weighted scores at 4.50, followed by Prisma Cloud at 4.40. The piece notes Forseti is deprecated and that Lacework is now Fortinet's FortiCNAPP, and flags diligence around Google's acquisition of Wiz. It is an editorial assessment, not a lab test, with pricing compared by model only.
A controlled study finds agent memory portability varies sharply: fixed-schema knowledge graphs survive model swaps while compressed notes degrade.
The study compares preserving an agent's history as raw long context, RAG chunks, compressed natural-language notes, or fixed-schema knowledge graphs across model upgrades, using 48 synthetic histories and two open-weight sub-10B-parameter models. Fixed-schema KG accuracy changed by only +0.0004 ± 0.0020 after a writer swap, while compressed NOTES shifted asymmetrically by +9.91 or -13.28 percentage points depending on migration direction. Mixed 50/50 embedding migrations captured only 4.96 of an 11.90-point RAG re-embedding gain; 80% of the NOTES deficit came from information lost at construction, and 81% of the RAG deficit from retrieval failures. Store-only repair of NOTES failed to reach 90% recovery in all 48 cases, while retaining raw histories enabled recovery in 34 of 48 for one direction.
A buying-guide comparison of twelve Kubernetes security vendors, from free OSS tools like Falco and Kubescape to paid platforms led by Sysdig and Aqua.
The article compares features and pricing mechanics for twelve Kubernetes security vendors including Sysdig, Wiz, Aqua Security, SUSE NeuVector, Snyk, Isovalent (Cisco), Palo Alto Prisma Cloud, Fairwinds, Red Hat ACS (StackRox), and Tigera. It argues CNCF-grade open source (Kubescape, Falco, Calico, NeuVector, Cilium/Tetragon) provides a free floor for posture, runtime, and network policy. Pricing hinges on node-versus-cluster-versus-developer billing units, and Cisco's Isovalent acquisition consolidates the eBPF layer inside a networking giant.
New tokeniser study shows search procedure, not optimisation objective, drives bits-per-byte performance across model sizes, vocabulary sizes, and multilingual settings.
The paper disentangles BPE and UnigramLM along two axes: optimisation objective (compression vs log-likelihood) and search procedure (bottom-up merging vs top-down pruning). Two new algorithms, BottomUpLL and TopDownComp, complete the 2x2 design space, and trained language models are evaluated on bits-per-byte and BLiMP across model sizes, vocabulary sizes, and English-only vs multilingual domains. Bottom-up tokenisers consistently achieve lower bits-per-byte in most settings, while BLiMP shows no consistent relationship with design choice.
Brinqa acquired PlexTrac to add pentest-driven validation of remediation, claiming the largest standalone unified exposure management vendor with 3,000+ customers.
Brinqa acquired PlexTrac to add offensive-security-driven validation that remediation actually worked, aiming to close the continuous threat exposure management loop from prioritization through verified fixes. Brinqa says the combination makes it the largest standalone vendor in unified exposure management, serving more than 3,000 customers across 57 countries, with PlexTrac continuing as a standalone product. Brinqa cited 164% year-over-year new-bookings growth in 2025 and inclusion in Gartner's inaugural Magic Quadrant for Exposure Assessment Platforms.
A controlled study of 60 deepfake detectors shows adversarial example transfer depends heavily on source-target compatibility, with source averaging understating vulnerability.
The study evaluates adversarial example transferability across 60 deepfake detectors spanning six backbones, two pretraining regimes, and five training-data configurations, using AutoAttack (AA) and Carlini-Wagner with Expectation over Transformation (CW-EOT). Transfer rises sharply when source and target share an exact backbone, architecture family, pretraining regime, or training data, with the dominant factor depending on the attack. Mean attack success rate is 7.21% under AA and 19.52% under CW-EOT for single sources, while a multi-source oracle reaches 64.48% after excluding exact matches, showing source averaging can substantially understate target vulnerability. The authors release 240,000 adversarially perturbed images, pairwise transfer results, detector configurations, and evaluation code.
Editorial scorecard ranks ten 2026 cloud detection and response platforms; Sysdig, Wiz, and CrowdStrike lead, with Wiz's Gem Security acquisition highlighted.
The editorial scorecard rates ten CDR platforms on real-time detection (30%), cloud telemetry depth, response automation, correlation, and value. Sysdig earns the best real-time detection score for its Falco- and eBPF-powered runtime telemetry, Wiz (8.7) folds acquired Gem Security's real-time CDR into its security graph, and CrowdStrike (8.7) leads response automation. Specialists Stream.Security, Skyhawk Security, Sweet Security, and the open-source Falco project are also assessed.
EvolveTrade lets LLM trading agents self-refine their tool-use policy from realized portfolio feedback, improving Sharpe ratios.
EvolveTrade treats a tool-using trading agent's system prompt as a text-parameterized policy that a Policy Agent revises after each update interval using accumulated decision traces and realized portfolio feedback, keeping the backbone LLM fixed. Experiments across multiple market regimes and two LLM backbones show improved Sharpe Ratio and Cumulative Return over fixed-policy baselines in most settings. Behavioral analyses show evolved policies increase code-mediated analysis and activate regime-relevant computations, with case-level attributions linking policy changes to returns.
Buyer's guide compares 12 MDR services, naming Huntress best value, CrowdStrike Falcon Complete for response authority and Expel for transparency.
The article compares 12 managed detection and response providers across response authority, tool bundling and pricing, highlighting Huntress for published SMB pricing and CrowdStrike Falcon Complete for unilateral containment. It stresses the consolidation landscape: Sophos completed its acquisition of Secureworks in February 2025 for approximately $859 million, and Arctic Wolf closed its purchase of BlackBerry's Cylance endpoint assets the same month. It also warns that only full-response contract tiers isolate hosts and kill processes, while lower tiers only triage or guide.
Separating design-direction exploration from code generation via structured specifications broadens UI alternatives without destabilizing output.
The paper proposes an inference architecture for vibe design agents that makes design direction an explicit intermediate decision: a Verbalized Sampling-inspired pre-pass proposes structured design specifications with typicality scores, an external selector samples one, and a downstream generator realizes it under fixed settings. Across 168 prompts with 1,255 paired comparisons per temperature, theme sampling broadens selection coverage and screenshot variation, with LLM-judge preferences varying across interventions and prompt complexity. An online experiment with over 300,000 tasks found the code-export increase statistically uncertain, though negative feedback events decreased alongside modest operational costs.
2026 enterprise browser guide ranks Island first and notes Mammoth Cyber's wind-down plus corrections to standard vendor shortlists.
An editorial guide assesses ten enterprise browser options, ranking category creator Island first for last-mile DLP and BYOD controls, followed by Palo Alto's Talon browser as a Prisma Access/SASE surface and Google Chrome Enterprise Premium for DLP on already-deployed browsers. It corrects common lists, noting SlashNext is browser-adjacent phishing and BEC defense rather than a managed browser, and that Mammoth Cyber has wound down independent operations. Microsoft Edge for Business is positioned as effectively free policy depth for Microsoft 365 estates, with Menlo Security offering an isolation-plus-browser blend.
Researchers introduce StudyBench, a physics benchmark showing self-evolution gains on textbook problems rarely transfer to olympiad-level questions.
StudyBench is a controlled physics benchmark splitting test data into an Application Set of difficult textbook problems and a Transfer Set of olympiad-level problems. Across three base models, representative self-evolution methods improved on the Application Set but rarely transferred to the harder Transfer Set. A guidance ablation reveals a Guidance Gap, and every method hits a Compute Plateau, indicating the remaining limits are method problems rather than data or compute problems.
Academics linked Chinese vendor Geedge Networks' Tiangou Secure Gateway source code to one of the Great Firewall's three traffic filtering capabilities.
US researchers presenting at USENIX Security reconstructed Geedge Networks' Tiangou Secure Gateway firmware from over 100,000 leaked files, including Git repositories with commit history, and matched its filtering behavior to sections of China's Great Firewall. They found only 1 of 3 characterized DNS injectors matched Geedge code, noted the system relies on memory-unsafe C components and copied third-party code, and said its bugs could aid future circumvention tools. Geedge also exports censorship tools to Kazakhstan, Ethiopia, Pakistan, and Myanmar. The newsletter additionally rounds up multiple breaches.
Learning to Coach trains a dedicated LLM coach to extract transferable experiential knowledge from a frozen actor's trajectories, beating self-refinement.
Learning to Coach (L2C) trains an LLM-as-a-Coach to extract actionable experiential knowledge from a frozen actor model's previous solution trajectories, optimizing rewards based on the actor's guided response correctness. It studies same-instance and cross-instance rewards, where cross-instance elicits knowledge that transfers to other problems. Across mathematical reasoning and interactive text-games, L2C outperforms self-refinement and untrained coaches, scales better with extra inference iterations than larger decoding budgets, and transfers to out-of-distribution tasks.
Trump signed an executive order declaring an emergency to bar foreign bulk-power equipment deemed a national security cyber risk.
The executive order, 'Declaring a National Energy Emergency to Secure the United States Bulk-Power System,' prohibits acquiring, importing, transferring, or installing foreign-produced bulk-power equipment and software deemed risky, citing fears of digital backdoors in Chinese-made grid gear. China supplies roughly 85% of solar supply chain capacity and is a major transformer manufacturer. The Energy Department has 120 days to develop implementing rules; the order revives a 2020 Trump-era measure the Biden administration had suspended after utilities found compliance difficult.
Paper proposes Generative Marketing Mix Modeling to causally estimate Generative Engine Optimization and Marketing effects on business outcomes.
The authors develop GMMM, a causal inference framework for measuring how often users see and notice a firm's name in generated answers, which standard marketing data ignore. For GEO it combines repeated generated answers with question counts, shares of generative-system usage and notice probabilities; for GEM it uses sponsored placement records with notice probabilities. The framework compares expected business responses under alternative treatment sequences, establishes identification conditions, and is evaluated on simulated product-recommendation answers in English and Japanese.
US offers up to $10 million for information on CyberAv3ngers' 'Mr. Soul,' whose IOControl malware targets US and worldwide ICS/SCADA devices.
The State Department announced a reward of up to $10 million for information on an online persona known as Mr. Soul (or Mr. Soll) tied to CyberAv3ngers, which conducts malicious cyber activity for Iran's IRGC Cyber-Electronic Command. CyberAv3ngers used the IOControl malware against ICS/SCADA devices in the US, Israel and worldwide, targeting products from Unitronics, D-Link, Hikvision and Baicells, according to Claroty and Armis analyses. The malware allows remote control of infected devices and lateral movement within victim networks. The reward follows an August action offering rewards for and sanctioning six Iranian government hackers and comes amid the widening Israel-Iran military conflict.
CanvasAnneal injects teacher reasoning traces into diffusion canvases during curriculum RL, improving diffusion LLMs on MATH500, Countdown, and Tau2.
CanvasAnneal is a curriculum-guided reinforcement learning framework for diffusion language models that addresses exploration bottlenecks in standard RL. It warm-starts exploration by injecting teacher-generated reasoning traces into the initial diffusion canvas, then gradually removes this guidance so the model generates reasoning trajectories independently. Across mathematical reasoning and tool-use benchmarks, it improves over standard diffu-GRPO on MATH500, Countdown, and Tau2 and accelerates reward improvement, though gains are task-dependent.
Proofpoint appointed Brian Levey as Chief Legal Officer and Puja Jaspal as Chief People Officer, strengthening its executive leadership team.
Proofpoint announced on September 8, 2026 the appointments of Brian Levey as Chief Legal Officer and Puja Jaspal as Chief People Officer. Levey previously served as Chief Business Affairs & Chief Legal Officer at Upwork and spent 13 years at eBay; Jaspal was Chief People Officer at Fastly with prior senior roles at Cisco, Visa, and Google. The hires support Proofpoint's stated expansion around human and AI-agent security.
The recent run of cyberattacks by in-development frontier models has got me thinking a lot about how our current incentive systems are not well suited for such fast technological transitions. The two primary power structures here are the rapidly growing technology companies and the federal government. The companies are incentivized to grow, so they can keep growing and keep scaling – in what is…
Researchers release CosmoH2G, a 6,189-episode hand-to-gripper dataset with a two-stage method for complex spatial robot manipulation.
The paper introduces a scalable acquisition pipeline using a handheld gripper to collect paired hand-gripper demonstrations, producing 6,189 episodes across 1,254 unique objects with higher spatial complexity than existing benchmarks. A two-stage framework first predicts sparse gripper keyframes (initial and terminal), then generates the full continuous action sequence conditioned on them, while learning gripper orientation and post-optimizing translation via grasping heuristics and kinematic consistency. Simulation and real-robot experiments show stable, precise hand-to-gripper transfer of complex spatial manipulations, outperforming traditional baselines.
Paper proposes a competence gate pooling language model forecasts with external priors, improving Brier score from 0.0771 to 0.0732 across 2,357 binary questions.
The paper defines a language model's relative competence as its marginal value beyond an available external forecast, and derives conditions under Brier loss where model disagreement improves that forecast. A competence gate estimates domain-level source weights from resolved outcomes, shrinks uncertain estimates toward a global weight, and recalibrates the pooled forecast. Across 2,357 resolved binary questions and five language models, it improves the external baseline from 0.0771 to 0.0732 Brier and beats global forecast combinations, though it defers to the market on ForecastBench. Across four Qwen models, verbal confidence failed to identify when the model outperformed the external forecast, while outcome-estimated competence supported better abstention.
AI-orchestrated campaign exploited PaperCut NG/MF RCE (CVE-2026-81578/82078), compromising 440+ instances at 395 organizations in 48 countries.
GreyNoise tracked a likely Russian-speaking actor using AI (OpenAI Codex harness plus a DeepSeek model) to develop, test, and deploy exploits for PaperCut NG/MF (CVE-2026-81578, CVE-2026-82078) starting 31 August 2026. The actor compromised at least 440 PaperCut instances across 395 organizations in 48 countries, achieving domain admin at 12 victims — fastest time to domain admin was five minutes and a US high school was fully compromised in seven minutes. Attack paths involved LSASS memory and registry secret harvesting, pass-the-hash to domain controllers, noPac attacks, account additions to Domain Admins, and DCSync to exfiltrate full NTDS.DIT credential dumps. Impact scope suggests access development potentially for handoff, with prior PaperCut intrusions historically leading to extortion.
Study shows visually grounded token embeddings in a small masked LM persist through training and improve object-property knowledge, but escape standard BabyLM benchmarks.
The paper implements ostensive definition for a small DeBERTa masked language model trained on 10M words, seeding visually grounded tokens with embeddings derived from labeled image regions before training. Visual initialization leaves a persistent, seed-replicated advantage on object-property knowledge (COMPS) and a corpus-tailored Visual-Property Swap benchmark covering color, material, size, and shape, but has no effect on most BabyLM grammar benchmarks. Synthetic grounding of previously unseeded words causally transfers the advantage to exactly those words.
Slovakia's NBU found an SMS-triggered backdoor in Russian-made NERO R-ONE traffic cameras, pausing a 279-unit deployment.
Slovakia's national security service NBU issued an alert against NERO R-ONE high-speed traffic cameras after finding a backdoor that grants shell and network access via SMS from hardcoded Russian phone numbers. The cameras are a rebranded version of the Russian CORDON PRO.M model by St. Petersburg firm Semicon, purchased via a Cyprus shell company under a €30 million EU-funded project. The report also found SecureBoot disabled, vulnerable web management, and unauthenticated live streams; the Interior Ministry paused deployment of 279 cameras pending independent assessment.
Drift-Constrained Optimization reformulates fine-tuning as update-direction selection, letting Qwen3 models improve target tasks within a behavioral drift budget.
The paper specifies a behavioral drift budget before optimization and shows that update direction is the remaining degree of freedom, reformulating fine-tuning as a direction-selection problem. In a stringent QA-only setting where instruct models must still generate multi-step reasoning at inference, a coarse layer-selective probe reverses the failure of QA-only fine-tuning. Across Qwen3-8B and Qwen3-14B, these directions substantially improve scientific reasoning and multilingual translation, matching or outperforming dedicated translation systems over 100+ languages and giving stronger initialization for reinforcement learning.
Microsoft's September 2026 Patch Tuesday fixes a record 964 CVEs, including two zero-days actively exploited in the wild.
Microsoft patched a record 964 CVEs in its September 2026 Patch Tuesday release, with 101 rated critical and 824 rated important, making it the largest Patch Tuesday to date. The release includes fixes for two zero-days that were exploited in the wild, with CVE-2026-81963 and CVE-2026-85880 highlighted in the release. Affected products span Windows core components, Office, Exchange, .NET, Visual Studio, Active Directory services (including AD CS and AD FS) and numerous Azure services. Tenable notes the release surpasses July's as the biggest ever.
Interpretability study traces how Qwen, Llama, and Gemma route query information and internal knowledge across layers when answering questions.
Researchers used layerwise interventions on hidden states to separate query-routing signals from target knowledge in language models, testing Qwen, Llama, and Gemma on country-continent questions with varied answer types. A pair-conditioned request direction strengthens before interventions alter downstream knowledge, opening a causal window while answer-supporting content is still forming. Trajectories differ by model: Gemma shows a partially overlapping mid-layer routing profile, while Llama has no sustained routing-effect window under the same gates.
Hackers stole over €35 million from 500+ French notary offices in a four-year BEC campaign; ANSSI spent two years helping evict the attackers.
A business email compromise campaign breached more than 500 French notary offices — about 7% of all French notaries per the Conseil Supérieur du Notariat — over four years, stealing more than €35 million by phishing initial access and silently modifying wire transfer details. France's cybersecurity agency ANSSI worked for two years behind the scenes to help notaries remove the persistent attackers, who had deep access; officials also feared hackers could issue fake notarized acts such as marriage certificates or forged real estate deals. No forged documents have been found so far, but notaries have added two-factor authentication and in-person requirements for banking details, and banks added extra checks in 2024. The newsletter also notes other incidents, including a $320 million Bitcoin extraction from Blockstream's Liquid Network and a JetBrains Cadence breach via TeamCity servers.
Audit of 22 frontier models finds widespread verbatim retrieval of published molecular property values, with higher reasoning increasing recall of memorized numbers.
An arXiv audit tests 22 frontier LLMs across 12 molecular regression benchmarks for verbatim retrieval of published values. More than 50% of the LLMs show verbatim retrieval on five datasets, and identical experiments are flagged 89% more often at a high reasoning level than at the lowest one. Suppressing retrieval moves model prediction errors closer together in relative terms, suggesting predictive capability is not determined solely by memorized values.
Microsoft details Teams IT-support impersonation campaigns deploying Node.js implants, AD reconnaissance, and WinRM lateral movement toward domain controllers.
Microsoft Threat Intelligence describes a human-operated campaign where attackers impersonate IT/helpdesk staff via Microsoft Teams external collaboration, talk users into granting remote sessions, and use RMM tools for interactive access. During the session they run PowerShell to silently install a malicious MSI that stages a portable Node.js runtime and obfuscated JavaScript implant for C2, executing follow-on payloads via rundll32. Operators then perform host and Active Directory reconnaissance, capture desktop screenshots, and pivot via WinRM on port 5985 to domain controllers and certificate authorities. The hands-on-keyboard chain, which can precede data theft and ransomware, blends into normal operations by relying on Teams, Quick Assist, msiexec, and Node.js; Microsoft shares hunting and mitigation guidance.
Gavel reads native skill-routing signals from a frozen LLM's forward passes with two linear maps, beating retrieve-and-rerank pipelines by up to 21.9 points on Qwen3-32B.
Gavel (Glance And Verdict from a frozen LLM) elicits skill routing from a frozen agent LLM using two trained linear maps that read mid-layer states, keeping all skill text out of context. A glance step scores the full library against compact per-skill banks built in one forward pass at installation; a verdict step resumes shortlisted skills' forward passes and fuses likelihood and yes/no judgments as a product of experts. It transfers zero-shot to three public benchmarks plus SkillTraj, a new benchmark of 372 simulated agent trajectories. On Qwen3-32B it beats progressive disclosure and retrieve-and-rerank pipelines adding 1.2B–16B external parameters by up to 13.4 points on written tasks and 21.9 when skills are needed mid-rollout.