Fetching from the wire…
01
02
03
04
05
06
07
08
Security2026-08-20 · source-backed
Reflex-Guard combines jailbreak-aware preprocessing, compact sentence-transformer embeddings and seven binary classifiers trained on 30,568 samples, reporting 95.9% recall end-to-end against 255ms for Llama Guard 2 and 723ms for SafeDecoding, with 100% detection of GCG suffix attacks and Base64-encoded prompts at defaults. (arXiv 2608.17556) Running locally also means the sensitive prompt never makes a round trip to a cloud safety API, which is the argument I'd lead with.
Each link below shares sources, entities, or timing with this story.
The method synthesizes security skills offline from known attacks and recorded agent failures, injects them into the system prompt at session start, and leaves them active through the tool-use loop. Across six models on RedCode the default all-classes skill dropped malware-gen...
Tencent's SkillJack work (arXiv 2608.03509) shows self-evolving agents launder malicious intent during skill extraction. Attack success rates of 56.2% on SkillX and 89.2% on Anything2Skill, with 80% of implanted skills surviving deletion of the original poisoned records. If yo...
arXiv 2607.15367 combines LLM planning with typed executors and adaptive error recovery. The typed-executor part is what I'd copy: constrain what the planner is allowed to emit rather than trusting free-form tool calls and validating after. Running the full stack locally is th...
arXiv 2607.05120 defines a new attack class: instead of hijacking what the agent does, corrupt which resources it acts upon, by injecting probabilistic delimiters that make untrusted data read as trusted metadata (resource IDs, data origin, tool-call history). On agents where...
MAFIA (arXiv 2608.03844) targets the two conditions that describe production and that prior attacks failed against: large benign memory pools and active input auditing. It adds placement strategy (probe memory, allocate injection budget, schedule writes to stay retrieval-compe...
arXiv 2608.13010 scores top-five retrieval candidates against ranks 6–20 of the same query to spot answer-anchor concentration, and separately compares documents to lexically distinct neighbors to catch coordinated density before any query arrives. Deployed jointly, attack suc...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.