Watermarking code by only coloring grammar-legal tokens, then turning up their temperature, because syntax was getting in the way.
trending115
01 02 Where Did the Repair First Go Wrong? Localizing the Origins of Silent Failures in Agentic Vulnerability Repair arxiv.orgWhere did the patch go off the rails, and can trace-based scoring pin the first security-divergent turn?03 OpenSSH 10.6 enables a post-quantum signature algorithm, so experimental keys need replacing helpnetsecurity.comOpenSSH 10.6 adds ssh-mldsa44-ed25519, and any experimental PQ keys you made earlier now get to be regenerated.04 Paper on agentic deanonymization of anonymized transcripts via web search, because masking is not a force field.05 Correct Verdicts, Flawed Reasoning: Structured Auditing of LLM-based Vulnerability Reasoning arxiv.orgPaper on auditing LLM vuln verdicts with structured logs, since correct answers still manage to hallucinate their way there.06 Semi-quantum secret sharing that claims collusion security against a quantum third party and a dishonest classical player.07 Paper on LLM router metadata leaking sensitive topics, with measurements and mitigations, because even the headers gossip.08 Semantic Behavioral Watermarking: Paraphrase-Robust and Forgery-Resistant Provenance for LLM Agents arxiv.orgAction-level watermarking for LLM agents, meant to survive paraphrases and make forgery need the key.09 Bounded Provisional Visibility: Controlling Poisoning Exposure in Continuously Ingested RAG Vector Stores arxiv.orgProvisional visibility for RAG vector stores, bounding poisoned content exposure instead of the usual ingest now, regret later.10 Can trigger tags survive an attacker who edits prompts or weights? This paper says, mostly not.11 Ethereum proposes native transaction assertions, with frame-transaction opcodes that let contracts inspect net state after execution.12 Constitution-guided watermarking tunes LLM watermarks to natural-language policy rules, with better post-paraphrase detection.13 Prompt injection gets a fenced-off proxy and a closed command set, then a red-team SOC eval tries to break it.14 P-VMI shows adversarial images can keep steering multimodal LLMs long after removal, thanks to the KV cache doing what caches do.15 OpenSSH 10.6 mostly says the quiet part out loud: AI found bugs, so they're shipping faster and changing some server defaults.16 Computes weakest trust assumptions from minimal attacks and stopping sets, not just the usual assume-honest ritual.17 Pump-and-Dump meets Honeypot Tokens: Detection and Analysis of Telegram Bait-and-Trap Schemes arxiv.orgPaper on Telegram-run pump-and-dump honeypots, with 3,677 cases and a pre-buy simulation detector, because rugs needed analytics.18 Can an LLM agent do traffic analysis? NetAgent says yes, using tool calls, replanning, and better OOD robustness.19 Three-track quantum healthcare forecast, with market sizing, evidence grades, and the usual urgent post-quantum security sermon.20 Paper on sequential cyber defense shows humans and DQN policies diverge late, while averaged scores conveniently hide it.21 Paper on backdooring sparse autoencoders to steer LLMs, including code insertion, because even the interpretability layer gets pwned.22 H-CRSPV bolts provenance checks onto LLM release pipelines, so omitted or misbound semantics get rejected instead of shipped.23 Multi-bit text watermarking with f-divergence budgets, via coding-based next-token biasing and the usual rate vs detectability trade.24 Paper on InterAcct, a cross-chain access-control scheme for permissioned chains, because leaking ACLs apparently needed a framework.25 Watermarking can make LLMs forget facts they were handed, and the paper tests six schemes plus fixes.26 How crypto is building identity, reputation and payment rails for an internet whose most active users may soon be software, and why the evidence still trails the hype. x.comWeb4 for agentic software: on-chain identity, reputation, and payments, with adoption numbers still trying to catch up to the pitch.27 How insecure are the off-the-shelf satellite radios we've been trusting? This paper audits them and finds firmware, protocol, and arch bugs.28 CorrectGuard tries to judge black-box guardrails without seeing the inputs, then compares ICL, embeddings, and finetuning across 13 datas...29 AgentSecGraph tracks security-sensitive deps, not just sinks, so it sees guards and source links sink-only analyses miss.30 MIRROR replicates LLM-agent messages over multiple routes and accepts them by majority digest, mostly killing MITM.