zknews

trending13

tag: llm-security ✕

01▲
Can Watermarking Techniques Help Prevent LLM Model Stealing? arxiv.org
Paper tests watermark-like logit perturbations against black-box LLM theft, then measures how much utility gets shaved off.
arxiv cryptography llm-security machine-learning-security model-stealing privacy watermarking
1 pt/dan_schnorr/4 hours ago/discuss
02▲
Toward Stronger Code Watermarking: A Grammar-Driven Approach to Optimizing the Trade-off Between Quality and Detectability arxiv.org
Grammar-driven code watermarking, using syntax-aware masking and role biasing instead of the usual token twitching.
arxiv code-generation llm-security machine-learning privacy watermarking
0 pts/mei/1 hour ago/discuss
03▲
PathMark: Protecting Intellectual Property of Mixture-of-Expert LLMs via Path Watermarks arxiv.org
PathMark watermarks MoE LLMs by hiding ownership in expert-routing paths, since dense-model tricks keep breaking.
arxiv llm-security ml-security privacy routing watermarking
4 pts/rkerr/7 days ago/2 comments
04▲
[tl;dr sec] #335 - Prompt Injection as Role Confusion, PHP Ecosystem Security, New MCP Spec tldrsec.com
Prompt injection, apparently, is role confusion, in a newsletter that also hits PHP hardening and the new MCP spec.
ai-security blockchain cloud-security cryptanalysis ecosystem-security linkedin llm-security mcp privacy prompt-injection supply-chain
2 pts/anyaw/9 days ago/discuss
05▲
Security--Fidelity Tradeoffs: The Hidden Cost of Prompt Injection Defense arxiv.org
Paper benchmarks prompt-injection defenses and finds the usual fix, stripping instructions, also breaks translation and editing; SecFid f...
arxiv evaluation llm-security prompt-injection robustness security
5 pts/ybauer/13 days ago/1 comment
06▲
On the Inseparability of Instructions and Data in Shared-Embedding Sequence Models arxiv.org
Formal impossibility result for prompt-injection defenses in shared-embedding models, where token reuse ruins the clean split.
arxiv formal-methods llm-security machine-learning prompt-injection security
0 pts/evanm/15 days ago/1 comment
07▲
ShareLock: A Stealthy Multi-Tool Threshold Poisoning Attack Against MCP arxiv.org
How do you poison MCP when each tool looks harmless? Split the payload with Shamir and let the model reassemble it.
arxiv llm-security privacy prompt-injection secret-sharing security shamir threshold-sharing
0 pts/ben/16 days ago/2 comments
08▲
Your System Prompt Is Not a Secret omsherikar.me
A writeup on system-prompt extraction attacks, scoring, and defenses, with leakage varying by model, because secrets are optional.
defense evaluation llm-security obfuscation privacy prompt-injection twitter
1 pt/ivan9/16 days ago/1 comment
09▲
Understanding and Mitigating Prompt Leaking Attacks in Real-World LLM-Based Applications arxiv.org
Paper on prompt leakage in 1,200 LLM deployments, finds >80% leak system prompts and proposes AREA, because of course they do.
arxiv cryptanalysis llm-security machine-learning-security privacy prompt-injection
0 pts/karl/19 days ago/3 comments
10▲
CodeSentinel: A Three-Layer Defense Against Indirect Prompt Injection in Code Contexts arxiv.org
The scary part is the repo, not the model, and CodeSentinel sanitizes code-context prompt injections before inference.
analysis arxiv code-llm llm-security prompt-injection sanitization security
0 pts/yuri_stderr/19 days ago/1 comment
11▲
SPARK: Security Knowledge Priming and Representation-Guided Knowledge Activation for LLM-based Secure Code Generation arxiv.org
Inference-time code hardening for LLMs: CWE prompts plus logit bias from safe/unsafe states, because retraining is so last season.
arxiv code-generation inference-time llm-security prompting retrieval secure-code-generation
1 pt/deadlock7/19 days ago/1 comment
12▲
Evaluating Prompting-Based Defenses Against Domain-Camouflaged Injection Attacks arxiv.org
Which prompting defense actually survives domain-camouflaged prompt injection? This paper benchmarks a few and mostly likes paraphrasing.
arxiv llm-security privacy prompt-injection security
0 pts/tara13/20 days ago/2 comments
13▲
Smarter Saboteurs, Better Fixers: Scaling & Security in Linear Multi-Agent Workflows arxiv.org
On HumanEval, bigger open-weight models are easier to hijack in linear multi-agent workflows, while a terminal fixer mostly undoes it.
adversarial-ml agentic-workflows arxiv jailbreaking llm-security multi-agent prompt-injection robustness
0 pts/ringlwe/1 month ago/3 comments