ZeroHour
DashboardNewsCVEsExploitsIncidentsAIBriefAnalyticsEntitiesWatchlistIOCsAsk the feedSourcesFeeds & APIKiosk mode

Search: “ap2”

29 stories

Searchkeywordsmart
DomainAllCyberAIWindow24h3d7d30dAllIn the wildImportant onlyShow duplicates
CategoryAllAI safety & security 2AI tools & infra 2Industry 1Phishing & fraud 1
SeverityAllcritical 0high 0medium 3low 1info 2

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

Signing the Transaction but Not the Decision: Whisper Attacks and a Binding Defense for AP2

Research shows AP2 agent-payment signatures can be manipulated into valid but wrong carts; proposed A-VIP defense binds signed intent to purchases.

A study demonstrates Whisper attacks on the AP2 agent payment protocol, where ordinary product-description text steers shopping agents into carts that pass every cryptographic check but no longer match user intent. Using Gemini Flash-Lite models specified by AP2's default sample agents, three attacks succeeded at 90%, 56%, and 73.3%, with the vulnerability spanning seventeen Google models, three agent frameworks, cross-vendor anchors, and Google's consumer assistant. The proposed A-VIP defense treats signed intent as a capability grant, binding credential lookups to sessions and cart lines to seen listings, blocking the first two attacks with zero false positives while surfacing unauthorized spending. The authors release A-VIP code, machine-checked invariants, and AP2-WhisperBench with 1,544 evaluation scenarios.

arXiv cs.CRupdated · 5d agofirst · 5d agoAI safety & security 2 sources1· 1 read

Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra

Cognition released SWE-2, a coding model post-trained from Kimi K3 that scores 50.0% on FrontierCode 1.1 Main, near Fable 5.1 at 64% lower cost.

Cognition introduced SWE-2, its most advanced coding model, post-trained from the 2.8T-parameter Kimi K3 base model. It achieves 50.0% on FrontierCode 1.1 Main, 73.0% on DeepSWE 1.1, and 92.8% on Terminal-Bench 2.1, beating Grok 4.6 and SWE-1.7 while matching Fable 5.1 and GPT-5.6 Sol at a fraction of the price. The company says it scaled reinforcement learning to the multi-trillion-parameter regime for the first time, using Pareto-informed cost penalties that train all reasoning-effort levels in a single run, tripled RL environments, and NVFP4/FP8 quantization-aware training. SWE-2 is available today in Devin Desktop and CLI, with rollout on Devin Web and Fusion.

Hacker News · AIupdated · 3d agofirst · 5d agoModel release 11 sourcesHN 58↑ · 15 comments1

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 · 8d agoAI research

openbmb/MiniCPM5-2B — new model trending #4 on Hugging Face

OpenBMB released MiniCPM5-2B, a dense 2B open-weights Transformer claiming 2B-class open-source SOTA for on-device deployment.

OpenBMB released MiniCPM5-2B, the second model in the MiniCPM5 series following MiniCPM5-1B, and it is trending #4 on Hugging Face. The dense 2B Transformer targets on-device, local, and resource-constrained deployment and claims 2B-class open-source SOTA while remaining competitive with 4B-class models. Reported strengths include coding, mathematics, long-context understanding, tool use, and agentic tasks; a tech report (arXiv 2506.07900), GitHub repo, and online demo accompany the release.

Hugging Face trending models · 10d agoModel release

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 · 7d agoAI research

LandingAI Releases Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity

LandingAI shipped Agentic Document Extraction Gen2 with DPT-3 Pro and DPT-3 Verity parsing models, adding usage-based billing, block-tree outputs, and word-level grounding.

LandingAI has generally released Agentic Document Extraction Gen2, rebuilt around two parsing models: DPT-3 Verity for deterministic transcription of digital documents with per-word bounding boxes and confidence scores, and DPT-3 Pro for layout-aware parsing of scans, handwriting, non-Latin scripts, and LaTeX math. Billing changes from a flat 3 credits per page to a page-plus-output-character model (Pro: 1 credit/page plus 0.5 credits per 1,000 output characters on priority; Verity: 0.3 plus 0.2), with an asynchronous standard tier at 0.5x price and vendor-claimed 25-80% cost reductions. Parse v2 returns a document-page-block tree with semantic IDs, normalized bounding boxes, and line- or word-level atomic grounding, replacing flat chunks; Gen1 client code will not run against Gen2 endpoints. Deployment options include US/EU cloud, VPCs on AWS, Azure, and Google Cloud, Snowflake, and air-gapped on-premises environments, with automated model routing planned for fall 2026.

MarkTechPost · 6d agoAI tools & infra

E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning

E2A-Bench, a 969-query financial chart reasoning benchmark, finds VLMs fail evidence-to-action consistency, with fine-tuning amplifying BUY:SELL bias 4-6x.

E2A-Bench is a 969-query benchmark built from 323 HS300 constituents across three input modalities with deterministic OHLCV-derived evidence anchors, evaluating grounding, reasoning-action consistency, evidence-confidence calibration, and directional coverage via UCR, RCI, ECI, and NDR metrics. Testing 20 VLMs showed the lowest-hallucination model ranked near the bottom on coverage with only 6.4% directional coverage, and oracle-aided verification reduced unsupported claims but could collapse coverage. Financial fine-tuning amplified the BUY:SELL ratio by factors of 4.21 to 4.68 across base-fine-tuned pairs.

Hugging Face daily papers · 3d agoAI research

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled autoregressive testbed shows validation losses must be analyzed per task, and image tokenizer choice affects joint multimodal text modeling.

Researchers built a pure-autoregressive testbed to study image tokenizers as the 'visual language' of unified multimodal models, tracking task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They found that losses exhibit distinct scaling behavior per task and rank tokenizers differently, and that I2T loss over a shared text vocabulary gives a more consistent loss–performance signal than T2I loss. Better reconstruction does not necessarily yield lower task-specific losses or stronger downstream performance, and tokenizer choice can affect text modeling under joint optimization. Case studies examine the discriminator, semantic supervision, and vocabulary size design axes.

Hugging Face daily papers · 8d agoAI research1

ReMoMask-2: Latent Retrieval-Augmented Masked Motion Generation

ReMoMask-2 rebuilds retrieval in the generator's latent space for text-to-motion generation, achieving lowest FID on KIT-ML and SnapMoGen.

ReMoMask-2 is a retrieval-augmented text-to-motion framework that constructs its retrieval database directly in the generator's pre-quantization latent space and aligns text queries through a distilled lightweight projector, eliminating the representation gap. The framework combines Hierarchical Bidirectional Momentum contrastive learning, Semantic Spatial-Temporal Attention, and Topology Structured Masking to handle hierarchical motion structure. The retriever achieves state-of-the-art accuracy, and ReMoMask-2 attains the lowest FID on KIT-ML and SnapMoGen, with a single mask-transformer stage outperforming the previous two-stage pipeline while delivering the fastest inference.

Hugging Face daily papers · 8d agoAI research

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled pure-autoregressive testbed shows task-specific validation losses rank image tokenizers differently, with I2T loss the most consistent signal.

Researchers built a controlled pure-autoregressive testbed and tracked task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They find losses should be analyzed per task because they exhibit distinct scaling behavior and rank tokenizers differently, and that the loss-performance relationship depends on the predicted token space. I2T loss, computed over a shared text vocabulary, correlates consistently with both generation and visual understanding performance after supervised finetuning. Case studies revisit the discriminator, semantic supervision, and vocabulary size as tokenizer design axes.

arXiv cs.AI / cs.LG / cs.CL · 7d agoAI research1

openbmb/MiniCPM5-2B-GGUF — new model trending #30 on Hugging Face

OpenBMB released MiniCPM5-2B, a dense 2B on-device model claiming open-source SOTA among 2B-class models.

OpenBMB released MiniCPM5-2B, the second model in the MiniCPM5 series following MiniCPM5-1B, as a dense 2B Transformer built for on-device and resource-constrained deployment with GGUF weights on Hugging Face. The team claims 2B-class open-source state-of-the-art performance, remaining competitive with 4B-class models in coding, mathematics, long-context understanding, tool use and agentic tasks. The release includes a tech report, GitHub repository and online demo, and is currently trending on Hugging Face.

Hugging Face trending models · 11d agoModel release1

Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models

An open methodology toolkit measures whether transformer language models contextualize fixed word forms across domains using bridge forms and layer-wise silhouette analysis.

The manual documents an open toolkit built around 'bridge forms' - identical written words recurring across two or more subject domains with a different sense in each - to test whether transformer language models individuate word occurrences by context beyond the embedding layer. It covers declarative specification of bridge forms, Wikipedia corpus acquisition, occurrence localization, layer-wise representation extraction, domain-pairwise silhouette measurement, and visualization, justifying each choice against failure modes such as sense contamination and subword-tokenization misalignment. It is a methodological and implementation reference and reports no empirical results.

arXiv cs.AI / cs.LG / cs.CL · 11d agoAI research

OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device

OpenBMB released MiniCPM5-2B, a 2.52B-parameter Apache 2.0 on-device model averaging 53.9 across 34 benchmarks, ahead of Qwen3.5-4B.

OpenBMB released MiniCPM5-2B, a 2,516,756,480-parameter dense LlamaForCausalLM model with grouped-query attention and a 131,072-token context, under Apache 2.0, runnable via vLLM, SGLang, llama.cpp, and Ollama. It averages 53.9 across 34 benchmarks versus 51.1 for Qwen3.5-4B, with strong tool-use (97.1 on tau2-Bench Telecom) and code results (69.1 LiveCodeBench v6, 46.4 SWE-bench Verified). Training combined 400B tokens of deep-thinking SFT, critic-based JustRL II RL teachers, and on-policy distillation merging 16 RL experts; datasets and intermediate checkpoints were published alongside the weights.

MarkTechPost · 8d agoModel release

From Protocols to Evidence: Bounded Claims for AI in Service of the Common Good

Paper proposes a rupture test and RISE AI architecture for evidence-bounded responsible-AI claims, framed via EU AI Act and NIST AI RMF.

The paper argues AI deployment intervenes in pre-existing institutional failures of responsiveness, belonging, care, and accountability, and must therefore evaluate both the system and the institutional rupture it enters. It reviews how the EU AI Act, NIST AI RMF, and ISO/IEC 42001 translate principles into protocols, and draws on Pope Leo XIV's Magnifica Humanitas to develop a rupture test linking institutional baselines to system evaluation. It distinguishes evidence-bounded deployment from measurement-bounded governance and introduces RISE AI, an architecture for bounded claims about Responsibility, Inclusivity, Safety, and Empowerment.

arXiv cs.AI / cs.LG / cs.CL · 5d agoAI policy

BreezeBlue/Breeze-TTS-2 — new model trending #19 on Hugging Face

BreezeBlue open-weights Breeze TTS 2, a bilingual text-to-speech model it ranks #1 among open-weight models on the Artificial Analysis TTS leaderboard.

BreezeBlue released open weights and Apache 2.0-licensed PyTorch inference code for Breeze TTS 2 on 2026-08-25. The text-to-speech model supports English and Chinese, voice cloning, reference-free voice design, voice direction, and inline vocal events like (laugh) and (sigh). Reported performance includes #1 open-weight ranking on the Artificial Analysis Elo leaderboard, under 40 ms time-to-first-audio, a 0.32 real-time factor on an NVIDIA H100, and about 7.7 GiB GPU memory for eager inference.

Hugging Face trending models · 21d agoModel release

Speculative Decoding in vLLM on AMD GPUs

vLLM benchmarks speculative decoding on AMD Instinct MI300X and MI355X GPUs across five drafting methods including EAGLE-3 and native MTP.

The vLLM project documents draft-and-verify speculative decoding support for AMD GPUs via ROCm, comparing native MTP, Gemma 4 MTP, EAGLE-3, DFlash, and DSpark drafting approaches. Output-token throughput effects varied with drafting method, proposal length, model family, draft checkpoint, workload, and acceptance behavior. The post also covers how to enable each method plus practical tuning and observability considerations.

Hacker News · AI · 9d agoAI tools & infraHN 141↑ · 52 comments1

YuE2 · Frontier Music with Symbolic Planning

YuE2, a 3.59B-parameter music generation model, scores 6.9632 on SongBench, beating Suno v5 via symbolic planning.

YuE2 is a music generation model of roughly 3.59B parameters and 28 layers supporting song creation, covering, and agentic editing through editable ABC symbolic scores. Its best-of-8 setting reaches 6.9632 on SongBench, the highest mean among 15 evaluated settings on WildSongBench (192 prompts), ahead of Suno v5 at 6.8721. The project also introduces MERT2, whose 632M-parameter encoders achieve state of the art on 14 of 15 MARBLE metrics, and SheetSage2, which transcribes beats, downbeats, key, chords, structure, and melody with SOTA on 10 of 13 benchmark metrics.

Hacker News · AIupdated · 4d agofirst · 5d agoModel release 6 sourcesHN 43↑ · 35 comments

Target leakage, not model class, explains reported accuracy in survey-based cardiovascular screening: a leakage-tiered audit of glass-box and tabular foundation models

Audits of 10 classifiers on BRFSS show target leakage, not model class, drives the reported 0.89 AUROC in survey-based cardiovascular screening.

The study benchmarks ten model classes, including glass-box and tabular foundation models, for prevalent myocardial infarction on 442,067 respondents of the 2022 BRFSS across five feature tiers of decreasing leakage risk. Removing two post-diagnostic features costs every model 0.049-0.051 AUROC and collapses performance into a 0.0045-wide band, and the explainable boosting machine matches all alternatives within 0.005 while scoring roughly 104x faster than the strongest foundation model. Frozen models transport within 0.002 AUROC to 2023 data; the authors conclude evaluation practice and feature sets, not model capacity, are the binding constraint.

arXiv cs.AI / cs.LG / cs.CL · 5d agoAI research1

Company Offering ‘100% Human-Written, Never AI’ Medical Research Is Entirely AI

404 Media found Research Gold, advertised as '100% human-written' medical research, is run entirely by AI with fake PhD staff.

Research Gold advertises PRISMA-compliant systematic reviews and meta-analyses drafted by PhD methodologists, but its listed experts are AI-generated personas that do not exist. Real methodologists, including evidence synthesis scientist Jenny Berrio, were listed without permission using photos copied from LinkedIn. Customer calls and quotes were handled by AI agents that insisted they were human, quoting $1,900 for a systematic review.

404 Media · Aug 11, 2026AI safety & security

CARDEA: Auditable Reasoning Grounded in Spatial Evidence for End-to-End Coronary Angiography Interpretation

CARDEA, a vision-language model trained only on public data, matches cardiologists on coronary angiography complexity assessment while exposing auditable bounding-box evidence.

CARDEA is a unified large vision-language model serving as the inference core of an end-to-end coronary angiography pipeline from multi-view videos to study-level diagnosis. It was trained on public datasets through visual alignment, self-distilled Chain-of-Box cold start, and reinforcement learning with verifiable rewards encouraging bounding-box reasoning. It reached 0.91 accuracy on dominance classification under domain shift and 0.90 on complexity assessment, comparable to two interventional cardiologists. RLVR raised zero-shot report generation vessel-severity macro-F1 from 0.513 to 0.686, while supervised imitation alone did not.

Hugging Face daily papers · 9d agoAI research

Claude's new system prompt really doesn't want to reproduce song lyrics

Anthropic published updated Claude consumer system prompts, including changes steering the model away from reproducing song lyrics, likely over copyright concerns.

Anthropic publishes system prompts for Claude.ai and Claude mobile apps, including historic revisions, and has reorganized them into an index with per-model pages such as the Haiku 4.5 page showing the original October 15, 2025 prompt and an updated January 18, 2026 version. The latest consumer prompt strongly discourages reproducing song lyrics, a behavioral constraint likely tied to copyright considerations. Prompts for Claude Cowork and Claude Code are not included in the published set.

Simon Willison · 13d agoAI safety & security

Quoting Rick Brewster

Paint.NET added a clean-room Direct2D rewrite for WINE, largely written by Anthropic's Claude and described as unreviewed 'vibe coded' code.

Rick Brewster says Paint.NET now ships a from-scratch, reverse-engineered Direct2D implementation (PaintDotNet.Windows.Direct2D1.Managed.dll) used under WINE via a /wine flag, since Direct2D was never completed well enough there. He credits the Claude coding assistant with writing most of the code, calling it largely 'vibe coded' and not thoroughly reviewed. Simon Willison shared the quote as an example of shipping AI-assisted systems code in production software.

Simon Willison · 14d agoAI tools & infra1

The OCUDU dApp Platform: An Open Runtime and E3 Interface for Real-Time AI-RAN

OCUDU open runtime lets third-party signed AI-RAN dApps run inside production 5G distributed units under three timing contracts, released as BSD-3 preview.

The OCUDU dApp platform provides an open runtime and E3 interface for executing signed AI-RAN applications inside a production 3GPP NR distributed unit, where prior dApp frameworks could only observe export streams. Modules run under three typed timing contracts: GPU receive-chain residency (Class A), the scheduler's 100 microsecond deadline (Class B), or non-blocking observer (Class C). On a GB10 gNB, dApps including an out-of-tree neural equalizer ran on a live cell without fallback. Platform, SDK, and a zero-hardware quickstart are public under BSD-3-Clause-Clear as a preview of the OCUDU AI-RAN Working Group 2.

arXiv cs.AI / cs.LG / cs.CL · 8d agoAI tools & infra1

ActionSplice: In-Flight Action Editing for Interactive World Models

ActionSplice enables in-flight action editing in chunk-autoregressive video world models via a lightweight corrector, avoiding rollback or waiting for the next chunk.

ActionSplice is an inference framework that formulates in-flight action editing for chunk-autoregressive video world models as Counterfactual State Transport (CST), where a lightweight corrector transports the interrupted backbone-native representation toward the matched state induced by the revised action. The world model and sampler remain frozen, and sampling resumes without replaying completed evaluations. Across minWM-Wan Action2V and HY-WM1.5, the retargeting variant CST-R reduces rollback-relative LPIPS by 61.5% and 75.9% versus direct condition swapping, while the temporal-splicing variant CST-T reduces suffix LPIPS by 56.1% and 77.5% with 2.73x and 1.69x pixel-ready speedups over waiting.

Hugging Face daily papers · 8d agoAI research

Cross-Model Agreement as a Deployment-Time Reliability Signal for Automatic Polyp Segmentation

Referee-Based Quality Estimation flags unreliable polyp segmentations at inference without ground truth, reaching ROC-AUC 0.960 with SegFormer-B0 referees.

RBQE measures agreement between a primary segmentation model and an independently trained referee on a 1,223-image external benchmark drawn from four public datasets. A cross-architecture SegFormer-B0 referee achieves the strongest signal (ROC-AUC 0.960), beating a Test-Time Augmentation baseline by 0.055 ROC-AUC under an identical protocol. Excluding trivially separable empty-mask cases, ROC-AUC falls to 0.876 (SegFormer-B0) and 0.783 (same-architecture control), but RBQE's margin over baselines widens. Progressive rejection of low-agreement predictions increases mean Dice of retained outputs, supporting selective prediction at the cost of one extra forward pass.

arXiv cs.AI / cs.LG / cs.CL · 6d agoAI research

Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence

Pelican-Sim 1.0 predicts future observations from visual context and robot actions; four-step autoregressive rollouts yield 5.67x speedup and raise policy success from 70% to 93%.

Pelican-Sim 1.0 is a general world model simulator for embodied intelligence that predicts future observations from visual context and robot actions using a 28-dimensional unified action space valid across heterogeneous embodiments. Sparse mixture-of-experts layers reduce FVD by 6.530 versus the dense backbone, and causal adaptation with few-step distillation yields a four-step autoregressive simulator achieving a 5.67-fold speedup over the 35-step model. Trained on roughly one million real-world and simulated trajectories, PSNR improves over the strongest baselines by 4.636 on AgiBotWorld Beta, 2.080 on RoboMIND, and 10.343 on RoboTwin. Downstream on RoboTwin, adding 500 generated trajectories to 50 demonstrations per task raises policy success from 70% to 93%, and policy evaluation reaches a Pearson correlation of 0.994.

Hugging Face daily papers · 6d agoAI research

Learning-Guided Planning in Large Dynamic Action Spaces: Budgeted Tree Search for One-to-Many Mobile Charging

LP-BTS uses graph proposal policies, learned critics, and budgeted PUCT search to plan mobile charging across dynamic action spaces up to 2,813 stops.

LP-BTS is a learning-guided planning architecture for one-to-many mobile charging, where N=250 sensors induce roughly 1,125 initial candidate charging stops. A graph proposal policy concentrates candidate support, a learned value critic evaluates leaves, and edge-budgeted PUCT compares simulated futures, letting a single frozen checkpoint cover action universes from 736 to 2,813 stops. On a sealed 30-scenario confirmatory bank it attains the highest observed survival (0.4545) and alive-AUC (0.8031), though its +0.0066 survival edge over the strongest engineered comparator is statistically unresolved.

arXiv cs.AI / cs.LG / cs.CL · 19h agoAI research

From Intent to Execution Grant: An Execution-Boundary Conformance Profile for High-Risk AI Actions

Researchers specify EBL-Core, an execution-boundary conformance profile binding AI agent intents, policies, and evidence into verifiable execution grants, validated with bounded tests.

The paper defines EBL-Core, a conformance profile deciding whether one fully materialized AI-generated candidate action may receive action-scoped execution authority. It binds a structured intent object, Root and Operational Policies, typed evidence, and a verifiable Decision Derivation through an Execution Release Contract, with lifecycle rules for Redemption and Revocation. Evaluation included 34 static vectors, 15 lifecycle checks, and 100 trials of 32 concurrent Redemption attempts yielding exactly one winner per trial. The authors state these bounded results demonstrate executability of the specified subset, not production readiness or complete mediation.

arXiv cs.CR · 5d agoAI safety & security1

PIDS-Bench: Evaluating Prompt-Injection Detectors Under Over-Defense, Obfuscation, and Distribution Shift

PIDS-Bench shows prompt-injection detectors scoring F1 above 0.98 still misclassify about one-third of external benign security-adjacent prompts, revealing provenance-sensitive over-defense.

PIDS-Bench is a frozen multi-axis benchmark that jointly evaluates prompt-injection detectors on attack detection and benign false-positive behavior at fixed thresholds, spanning in-distribution inputs, hard-benign prompts, obfuscated attacks, and domain/structural distribution shifts. It evaluates seven detectors plus a rule-based lower-bound reference. A detector exceeding F1 = 0.98 on held-out data still misclassifies roughly one-third of an externally-sourced benign security-adjacent subset, and no internal detector reaches F1 >= 0.95 with hard-benign FPR <= 0.10 on the stress distribution. Hard-negative augmentation nearly eliminates over-defense on curated stress inputs but leaves it intact on externally-sourced prompts, a pattern termed provenance-sensitive over-defense.

arXiv cs.CR · 2d agoAI safety & security