Search: “Reco”

The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

Position paper defines recursive self-improvement for AI, introduces the Headroom-Closed Index and an autonomy roadmap toward genuine recursive meta-improvement.

The paper uses the Headroom-Closed Index to diagnose limitations of existing LLMs and frames recursive self-improvement (RSI) as a staged roadmap: improvement-execution, improvement-strategy, experience-acquisition, and environment-adaptation autonomy, culminating in recursive meta-improvement. It examines RSI across scientific discovery, embodied intelligence, and software engineering, highlighting differing requirements and development speeds. Drawing on industry practices and preliminary empirical evidence, it connects RSI research with practical systems and identifies key challenges to achieving genuine RSI.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:44 UTC · 5d agoAI research 2 sources1

ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents

ScienceBuddy couples harness evolution with model reinforcement learning so scientific agents continually self-improve from researcher feedback in an interactive workspace.

The authors release ScienceBuddy, an interactive scientific research workspace that transforms researcher requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. Its recursive-in-recursive self-improvement paradigm couples harness evolution (inner recursion, model fixed) with model reinforcement learning under the improved harness (outer recursion). Case studies span four scientific task families, and the system is released to the scientific community as a research product.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 17:55 UTC · 17h agoAI research 2 sources

A stranger has been reading Salesforce and ServiceNow portals worldwide for 17 months

Reco tracks City-Forum, an actor abusing Salesforce and ServiceNow guest-user access from one IP for 17 months to enumerate portal data.

Reco's City-Forum investigation found one IP address, tied to a dormant 2002 domain, anonymously enumerating Salesforce Experience Cloud and ServiceNow portals across telecoms, banks, enterprise software vendors, and public sector since at least March 2025. No vulnerability was exploited and no credentials were used; the actor abused over-permissioned guest users, including a previously undocumented ServiceNow portal search endpoint, with the busiest single target logging over 560,000 events. Audit logs show what was requested but not which records or fields were returned, so defenders must replay anonymous requests to assess exposure.

Help Net Security·Aug 12, 00:00 UTC · Aug 12, 2026Threat actor

ReCite: Agentic Reasoning for Faithful Citation

ReCite is an agentic citation framework using claim-level reasoning and verification, outperforming large generative models in strict citation accuracy.

ReCite is a decoupled agentic framework that orchestrates location perception, intent-aware query planning, and reflective verification for citation recommendation. Trained on synthesized reasoning trajectories, the agent verifies claim-evidence consistency and triggers self-correction loops when retrieved candidates lack logical support. Experiments show the lightweight framework outperforms state-of-the-art massive generative models in strict citation accuracy, addressing misattribution where cited papers are real but logically unsupportive.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:59 UTC · 7d agoAI research

One Attacker Has Scraped Both Salesforce and ServiceNow Portals Since 2025

Reco tracks the City Forum campaign: one Contabo VPS has scraped Salesforce and ServiceNow guest portals across finance, telecom and public sector since March 2025.

Reco researchers identified infrastructure at 158.220.87.79, a Contabo VPS using Go's default net/http user agent, that has harvested records from Salesforce Experience Cloud and ServiceNow customer portals across telecom, financial services, enterprise software and public sector targets for over a year. The tool abuses over-privileged persistent guest users, enumerating Salesforce Aura sites (560,000+ events at one target), probing Lightning Web Runtime via the UI-API across API versions v56.0-v66.0, and hitting the poorly documented ServiceNow Service Portal search endpoint. No named victims or actor attribution yet, and activity is ongoing with rising volume. Defenders are advised to hunt the Go-http-client user agent and specific IP in Salesforce Event Monitoring logs and syslog_transaction on ServiceNow, then tighten guest sharing rules and Knowledge Base read criteria.

The Hacker News·Aug 18, 11:30 UTC · 28d agoThreat actor in the wild1

Not All Relations Are Equal: Relation-Balanced and Calibrated Graph Learning for Provenance-Based Intrusion Detection

RECAL improves provenance-based APT detection with relation-balanced masked graph learning and calibrated errors, reaching 99.99% F1 on DARPA E3 datasets.

RECAL is an unsupervised framework for provenance-based intrusion detection that uses relation-balanced masked graph learning to capture rare interaction patterns, addressing statistical heterogeneity where relation frequencies differ by roughly 140,000X in CADETS. It calibrates reconstruction errors against each relation's benign error distribution to produce comparable anomaly evidence and reduce false alarms. On three DARPA E3 datasets, RECAL achieves F1 scores of 99.99%, 99.93%, and 99.99%, outperforming the best baseline on each dataset, and reduces mean false positive rate by approximately 105X, 4X, and 41X versus the lowest-FPR baseline.

arXiv cs.CR·Sep 15, 00:31 UTC · 1d agoResearch

Four in Five AI Tools Run with No IT Oversight, New Research Finds

Reco's research finds four in five AI tools run without IT oversight, linking expanding shadow AI usage to a surge in vulnerability disclosures.

Reco's new report states that roughly 80% of AI tools in organizations operate with no IT or security oversight. The research connects growing shadow AI adoption to an increasing number of vulnerability disclosures. The findings highlight governance gaps created by employees deploying unsanctioned AI services.

Infosecurity Magazine·Aug 26, 13:00 UTC · 20d agoAI safety & security

Google Search Makes It Harder to See Where a Link Really Goes Before You Click

Google now routes some Search results through opaque google.com/goto redirects, weakening hover-to-verify anti-phishing checks.

Google has begun serving some organic search results as opaque google.com/goto?url= redirects whose destinations can only be resolved server-side by Google, likely to raise scraping costs for rank trackers and archival services. The change removes the pre-click hover preview of the true destination URL, undermining a long-standing anti-phishing habit for spotting lookalike, typosquatted, or search-optimized phishing domains. Security teams are advised to rely on layered defenses such as domain reputation, DNS and web filtering, browser isolation, and user training rather than hover text.

GBHackers·Sep 15, 13:42 UTC · 21h agoIndustry

ThreatsDay: GhostJacking AI Attacks, EtherHiding ClickFix, Cursor CLI Flaw + 17 More Stories

A weekly bulletin aggregating short security updates, including the City-Forum data-theft campaign, a ShipMonk breach, a Cursor CLI flaw, and GhostJacking AI attacks.

The Hacker News ThreatsDay Bulletin bundles roughly 20 short updates across cloud services, AI tools, malware, breaches, and scams. Highlights include the City-Forum campaign pulling data from unauthenticated guest access in Salesforce Experience Cloud and ServiceNow Service Portals since March 2025, and a ShipMonk breach exposing Trezor customer order data for orders in seven countries between May 10 and August 8, 2026. Other items cover a patched Cursor CLI flaw that let cloned repositories run commands before the workspace-trust prompt, Okta's analysis of the Work Panel vishing console used by actors like UNC6671, and GhostJacking AI agent hijacking via a patched Claude Desktop sandbox escape. Meta also launched an on-device WhatsApp Scam Alert machine learning model that keeps message content on the device.

The Hacker News·Aug 18, 16:30 UTC · 28d agoIndustry

ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation

ActReview post-trains Qwen3-8B-Base on OpenReview rebuttals to generate actionable peer-review feedback with grounded revision suggestions, benchmarked on 1,000 curated instances.

The paper defines Actionable Peer-review Generation as diagnostic claim generation plus revision suggestion generation and introduces ActReview, a rebuttal-guided post-training framework. From OpenReview review-rebuttal threads the authors build ActReview-40K, aligning reviewer weaknesses with author responses grounded in localized paper evidence, and post-train Qwen3-8B-Base with multi-task SFT followed by GRPO using weakness-specific rubric rewards. They also release ActReview-Bench, a human-curated 1,000-instance benchmark, on which ActReview outperforms prior specialized review-generation models on actionability and grounding while remaining competitive with strong prompt-based LLMs. Human evaluation confirms improved revision usefulness but identifies a remaining gap in technical accuracy.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

Week in review: Salesforce and ServiceNow portals exposed for 17 months, exploited Metabase 0-day

Weekly digest: exploited Metabase zero-day breached Framework; Salesforce/ServiceNow portals read for 17 months; Microsoft patched 400+ flaws.

Help Net Security's week in review aggregates top stories: a 'City-Forum' campaign tracked by Reco has been pulling records from Salesforce and ServiceNow portals worldwide for 17 months, and Framework suffered a breach via an exploited Metabase zero-day exposing customer contact and IP data. It also covers Microsoft's August 2026 Patch Tuesday fixing 400+ flaws including exploited zero-day CVE-2026-68820, Cisco's fix for exploited firewall DoS bug CVE-2026-20349 (added to CISA KEV), and a second N-able N-central hotfix for actively exploited CVE-2026-18577. Other items include GitHub expanding Dependabot malware alerts to eight package ecosystems and EU AI Act enforcement beginning on 2 August 2026.

Help Net Security·Aug 16, 00:00 UTC · Aug 16, 2026Industry in the wildCVE-2026-18577CVE-2026-68820CVE-2026-203491

DriveZero: End-to-End Driving Beyond Human Demonstrations

DriveZero pairs a frozen vision-foundation-model perception stack with a PPO-trained closed-loop RL teacher to beat replay experts on nuPlan.

DriveZero is an end-to-end camera-only autonomous-driving planner that separates perception and action. Its DriveVFM perception backbone consolidates frozen vision foundation models (DINOv3, SigLIP2, SAM, Depth Anything V2) from raw images without task annotations, while DriveRL trains a privileged PPO teacher policy through closed-loop rollouts in interactive worlds built from real driving logs. The planner distills this teacher, achieving a 93.57 mean nuPlan score across Val14, Test14-hard and Test14-random splits and beating the Log-Replay expert on all three. It also sets state of the art on NAVSIMv1, NAVSIMv2 and closed-loop HUGSIM without human trajectory supervision.

Hugging Face daily papers·Sep 4, 20:00 UTC · 11d agoAI research

Cybersecurity jobs available right now: July 7, 2026

Help Net Security·Aug 7, 10:57 UTC · Aug 7, 2026Vulnerability

5 useful things you'll learn in my new post-training textbook (shipping now!)

Nathan Lambert's new RLHF and post-training LLM textbook covers PPO, GRPO, GSPO, CISPO and related techniques, freely available online.

Nathan Lambert's book 'Reinforcement Learning from Human Feedback: Aligning and Post-training LLMs' is now shipping from Manning. It covers policy-gradient algorithms including PPO, GRPO, GSPO, CISPO, and RLOO, plus loss aggregation, truncated importance sampling, asynchronous RL systems, and post-training topics like rejection sampling, outcome reward models, and on-policy distillation. The book is freely available online with a 12-hour course, codebase, and exercises.

Interconnects·Aug 10, 13:02 UTC · Aug 10, 2026AI research

ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation

ActReview post-trains Qwen3-8B-Base on 40K rebuttal-derived instances with rubric rewards to generate actionable, grounded peer-review feedback, plus a 1,000-instance benchmark.

The framework builds ActReview-40K from real OpenReview review-rebuttal threads, aligning reviewer weaknesses with author responses and grounding feedback in localized paper evidence. Qwen3-8B-Base is post-trained with multi-task supervised fine-tuning followed by GRPO using candidate-aware, weakness-specific rubric rewards. Experiments show improved actionability and grounding over prior specialized review-generation models, supported by ActReview-Bench, a human-curated 1,000-instance evaluation set. Human evaluation confirms better revision usefulness while noting a remaining gap in technical accuracy.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:30 UTC · 7d agoAI research

Google’s new search redirects make links harder to check before you click

Google routes some search result links through encoded google.com/goto?url= redirects, breaking hover-to-verify link checks and raising scraping costs.

Google has started routing some search result links through opaque google.com/goto?url=... redirects using a custom Google-specific encoding, with the destination visible only via the redirect response's Location header. Google confirmed the rollout as an anti-abuse measure, most likely to make bulk extraction of destination URLs from search results more difficult and costly. Malwarebytes warns the change undermines the standard hover-before-clicking safety advice, while legitimate rank-tracking, SEO auditing, archival, and accessibility tools now face the same rate limits and costs as abusive scrapers.

Malwarebytes Labs·Sep 14, 14:17 UTC · 1d agoIndustry

A Princeton Researcher Proposes Recurrent Looped Transformer (RLT) that Carries Decoder State across Every Token, Fixing 96 Blocks per Token with Unbounded Temporal Depth

Princeton researcher Yifan Zhang proposes Recurrent Looped Transformer, carrying full decoder state across every token for unbounded temporal depth.

Yifan Zhang's technical report defines the Recurrent Looped Transformer (RLT), pairing a causal encoder with a recurrent decoder whose final output and layerwise sliding-window attention cache carry into every subsequent token with no prompt-response boundary reset. The reference configuration ties 48 encoder and 48 decoder layers, executing 96 logical blocks per token while the state path grows to 48t blocks after t tokens at fixed per-token compute. The report details RL replay contracts that rebuild all states under current parameters and exact prefix snapshots for multi-turn serving, but explicitly reports no measured efficiency, reasoning quality, or scaling results.

MarkTechPost·Sep 13, 17:02 UTC · 2d agoAI research1

Google’s New Search Redirects Make It Harder to Check Where Links Lead Before Clicking

Malwarebytes reports Google's new encoded google.com/goto?url= redirects break hover-preview link checking, weakening a common phishing defense.

Google now routes some search results through opaque google.com/goto?url= redirects using custom encoding, so browser link previews no longer reveal the true destination, only the claimed label above the result. Malwarebytes found the final destination is visible only in the redirect response's Location header, complicating hover-based safety checks as well as scraping, archiving, and audit tools. The change arrives amid malvertising, search-result poisoning, and fake installer campaigns like the recent Node.js infostealer lure. Google says it deploys measures against evolving abuse but did not explain the change.

Cyber Security News·Sep 15, 12:21 UTC · 22h agoResearch

How to build an exposure management program the business trusts: Lessons from Tenable’s CSO

Tenable's CSO describes an AI-driven exposure management program that consolidates tool sprawl and translates cyber risk for boards.

A Tenable blog post shares lessons from CSO Robert Huber on moving to an AI-driven exposure management program. It argues tool sprawl and data silos hinder holistic risk assessment and that exposure management unifies attack-surface data into business-level metrics for the C-suite.

Tenable Blog·Aug 27, 14:30 UTC · 19d agoIndustry

When the prompt becomes the payload: A practical pen-testing guide for GenAI, LLM and RAG applications

CSO Online publishes a practical penetration-testing guide for GenAI, LLM, and RAG applications, covering prompt injection, retrieval poisoning, and tenant isolation testing.

The guide frames LLM applications as attack graphs spanning prompts, retrieval layers, vector stores, tools, identities, and downstream APIs, arguing that conventional web testing misses instruction-vs-data channel risks. It builds on OWASP prompt injection guidance (direct vs. indirect injection) and NIST's 2025 adversarial machine-learning taxonomy, noting that RAG and fine-tuning do not remove injection risk. Recommended practices include documenting trust transitions across components, using canaries and synthetic records to avoid test side effects, running multi-turn and obfuscated injection campaigns, and verifying chains from poisoned documents to observable state changes. It also details testing RAG pipelines via controlled document poisoning across metadata, OCR layers, and code comments, plus cross-tenant isolation checks on retrieved document IDs.

CSO Online·Sep 9, 10:00 UTC · 7d agoAI safety & security1

Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps

Hugging Face guide fine-tunes a 350M-parameter model with 100 GRPO steps to improve structured output reliability.

A Hugging Face blog post demonstrates fine-tuning a 350M-parameter model using GRPO (Group Relative Policy Optimization) with TRL over 100 training steps. The stated goal is more reliable structured outputs from small language models. No article body was available, so details beyond the title are limited.

Hugging Face Blog·Sep 3, 00:00 UTC · 13d agoAI tools & infra

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

NeoHorse-1 introduces agentic post-training with intelligent routing that lifts agent benchmark scores at 4B and 9B scales, prototyping recursive self-improvement.

NeoHorse-1 is a family of agent-native models trained through agentic post-training: routing-harness logs (predicted capability demand, service tier, interaction) become structurally validated training data organized into a three-stage SFT curriculum plus routing-guided on-policy distillation. Capability-guided allocation converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop. Post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B across eleven agent, tool-use, coding, and instruction-following benchmarks. The authors position it as a prototype of harness-mediated recursive self-improvement.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

The Frontier AEO Tracker: What Astra Chooses (and every other frontier model, and what you can do about it)

Latent Space launches an AEO tracker scoring 7 frontier models' product recommendations across 161 categories, revealing generational bias flips.

Latent Space built a tracker measuring Answer Engine Optimization by running 6 prompt variations across 7 frontier models with search enabled over 161 product categories, scoring first choices, alternatives, mentions, and anti-recommendations. It found 28 categories with a universally dominant primary choice and observed soft biases, such as models favoring their own lab's coding agents. Analysis of Anthropic's Sol→Astra and Opus→Fable generations showed newer models consulting fewer sources and being less likely to change answers when questions are paraphrased.

Latent Space·Sep 7, 21:32 UTC · 8d agoAI research

ThinkPrior: Zero-Rollout Difficulty Priors for Cold-Start Prompt Selection in RLVR

ThinkPrior builds zero-rollout difficulty priors via an offline verifier-anchored pass, halving silent groups in RLVR and cutting wasted rollouts on Qwen2.5-Math-7B.

In GRPO-based RLVR, groups where all rollouts are correct or all are wrong yield zero advantages and consume about 39% of a run's rollouts under uniform sampling. ThinkPrior initializes a Beta posterior from an external anchor pass's verifier-scored pass rate, selecting prompts by expected learnability before any target-policy rollout, without changing the loss or optimizer. On Qwen2.5-Math-7B across sixteen seeds it more than halves early silent groups and cuts wasted rollouts through step 30 by nearly a fifth, with no detected final-accuracy difference. The ThinkPrior+DAPO composition reduces generated rollouts by 10.6% at an equal 3,840-rollout update budget.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:30 UTC · 7d agoAI research1

ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement

Researchers propose ModularRSI, a modular benchmark-disjoint recursive self-improvement framework that evolves agent harnesses across five modules, improving TB2.0 and SWE-Bench Verified results.

ModularRSI targets generalizable recursive self-improvement (RSI) for agent harnesses by contrasting successful and failed trajectories for the same task and aggregating evidence across tasks to find recurring behavioral deficiencies. It decomposes the evolvable harness into five modules—Agent Loop, Tool Use, Observation Management, Context Management, and Task Completion Detection—each evolved independently within a restricted scope, then integrated with conflict resolution. Using 2,000 executable evolution tasks disjoint from evaluation benchmarks, it shows consistent gains on TB2.0 and SWE-Bench Verified and transfers across different foundation models.

Hugging Face daily papers·Sep 13, 20:00 UTC · 2d agoAI research

Rust Supply Chain Attack on arrayref: Significant Overlap with DPRK Campaigns

Malicious versions of the Rust crate arrayref executed a compile-time backdoor, with infrastructure overlapping recent DPRK supply chain attacks on Mastra and axios.

Wiz researchers report that malicious versions of the arrayref Rust crate ran a backdoor during compilation, compromising developer build pipelines. The campaign's command-and-control infrastructure significantly overlaps with infrastructure used in recent DPRK supply chain attacks, including campaigns against Mastra and axios. Developers using affected crate versions should audit their builds and review dependencies.

Wiz Blog·Aug 20, 15:56 UTC · 26d agoThreat actor in the wild1

Opaque recurrence, and other AI terms that you should probably know

TechCrunch updates its plain-English glossary defining common AI terms from AGI and agents to chain-of-thought reasoning.

TechCrunch maintains a regularly updated glossary of AI terminology, defining terms such as AGI, AI agents, API endpoints, chain of thought, coding agents, compute, deep learning, and diffusion. It highlights 'opaque recurrence', the reasoning technique in OpenAI's new Astra model that has drawn attention from AI safety researchers. The piece is an educational living document rather than new research or a product announcement.

TechCrunch · AI·Sep 7, 19:24 UTC · 8d agoAI industry

VeriScene: Reconstructing Crime Scenes from Legal Evidence via World-Model Agent

Researchers present VeriScene, a world-model agent that reconstructs crime scenes from forensic photos and witness statements with traceable, physically plausible output.

VeriScene orchestrates a world model to fuse forensic photographs and witness statements of varying reliability into cited narratives and physically plausible re-enactment videos. On a 25-scenario benchmark with planted unreliable testimony, it reaches 0.9014 evidence coverage and 0.7217 factual consistency on 20 test scenes. It outperforms an end-to-end multimodal-LLM baseline by 20.35% in factual consistency and 34.88% in temporal coherence at USD 1.82 per scene.

arXiv cs.CR·Sep 8, 07:14 UTC · 8d agoAI research

New Ted Backdoor Hides Inside Victims' Own HAProxy Builds to Intercept Web Traffic

Rapid7 found a new backdoor, ted, compiled into trojanized HAProxy at two South Korean organizations, with medium-confidence attribution to North Korean actors.

Rapid7 documented a previously undocumented Linux toolkit named ted compiled into the HAProxy load balancer binaries of two South Korean organizations in the automotive and media sectors. The implant intercepts web traffic, serves altered pages only to filtered visitors, and hides C2 exchanges from backend logs and HAProxy statistics; a companion RAT, curlRAT, beacons on a default 12-hour schedule. The toolkit also trojanizes crond, sshd, agetty, atd, and polkitd binaries and sanitizes logs and bash history. Rapid7 attributes the activity with medium confidence to North Korean state-sponsored actors, with domain infrastructure overlapping APT37 listings in maltrail and delivery resembling the Operation SyncHole campaign.

The Hacker News·Sep 4, 14:51 UTC · 11d agoThreat actor in the wild

RetroThinker: Enabling Retrospective Thinking in Speech LLMs

RetroThinker is a post-training framework letting the Moshi speech LLM self-correct reasoning mid-stream, adding 11% GSM8K accuracy at similar latency.

Researchers introduce RetroThinker, a multi-stage post-training framework that equips the Moshi speech LLM to verify and forward-correct chain-of-thought steps during streaming inference. It combines supervised fine-tuning on curated retrospective thinking data with length-based direct preference optimization (DPO). On GSM8K it achieves an 11% absolute accuracy gain over non-retrospective baselines at comparable latency.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:41 UTC · 5d agoAI research1

RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments

RSIAgent, a training-free multi-agent framework, builds reusable environment memory enabling Kimi-K3 and GLM-5.3 to beat GPT-6.

RSIAgent is a training-free framework for recursive self-improvement through autonomous memory construction, coordinating curriculum, actor, and verifier agents. It uses broad-then-deep exploration to capture environment structures, hidden constraints, and causal dependencies, and freezes the resulting memory for direct reuse without parameter updates. On OSWorld-v2 and Agent's Last Exam it substantially improves strong open-source models, enabling Kimi-K3 and GLM-5.3 to outperform frontier closed-source models including GPT-6.

Hugging Face daily papers·Sep 13, 20:00 UTC · 2d agoAI research2

Bad Likert Judge: A Novel Multi-Turn Technique to Jailbreak LLMs by Misusing Their Evaluation Capability

Unit 42 details the Bad Likert Judge multi-turn jailbreak that abuses LLMs' evaluation capability, raising attack success rates over 60% across six frontier models.

Palo Alto Networks Unit 42 describes the Bad Likert Judge technique, a multi-turn jailbreak that asks a target LLM to act as a Likert-scale judge scoring the harmfulness of example responses. The highest-rated example in each scale can carry harmful content, bypassing the model's internal guardrails. Testing across six state-of-the-art text-generation LLMs showed an average attack success rate increase of more than 60% versus plain attack prompts, with tested models anonymized. The technique targets edge cases rather than typical use, and the article positions the work as guidance for defenders on potential jailbreak risks.

Palo Alto Unit 42·Aug 17, 12:34 UTC · 29d agoAI safety & security

Inoculation Midtraining with Learned Neologisms

Inoculation Midtraining confines unsafe LLM behavior to a neologism-marked context, reducing misalignment after unsafe post-training but leaking under nearby contextual cues.

The paper introduces Inoculation Midtraining, which teaches a base model during midtraining that unsafe behavior belongs to a context marked by a learned neologism token, then post-trains on unsafe data within that context. Across supervised fine-tuning and RL post-training regimes, the technique reduces misalignment while preserving transfer of benign properties like German or Shakespearean prose. However, it does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. The authors conclude it is not yet a load-bearing component of a developer safety framework.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:12 UTC · 1d agoAI safety & security

Context Engineering Inside the Harness: 4 Mechanisms That Beat Context Overflow and Goal Loss on Long-Horizon Tasks

Survey of four harness mechanisms—context budgeting, compaction, todo-state, and memory—that keep long-horizon LLM agents on task across 200+ tool calls.

The article details how agent harnesses, not larger context windows, solve context overflow and goal loss on long-horizon tasks, citing Chroma's Context Rot report showing 18 LLMs (GPT-4.1, Claude 4, Gemini 2.5, Qwen3) degrade on long inputs. Concrete implementations include LangChain Deep Agents offloading tool responses over 20,000 tokens to the filesystem and truncating old tool calls at 85% window usage, and Claude Code capping auto memory at 25KB while re-reading the 5 most recently modified files after compaction. OpenAI's Responses API now offers server-side compaction via context_management with a standalone /responses/compact endpoint, which Codex uses for long-running coding tasks. Manus reports a roughly 100:1 input-to-output token ratio per ~50-tool-call task, motivating todo.md state recitation to prevent goal drift.

MarkTechPost·Sep 13, 05:56 UTC · 3d agoAI research1