Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-24 · high
Stricter privacy instructions cut the leak rate but never eliminated it, across DeepSeek and Claude agent setups.
Why now: The paper posted to arXiv in August 2026, giving teams already running tool-calling agents a concrete leak-rate range to test their own systems against.
Researchers found AI agents disclosing protected user data in tool-call arguments in 20.8% to 75% of trials despite privacy instructions, per a study testing DeepSeek and Claude agents. Any team gating tool-call arguments on nothing but a system-prompt privacy policy is exposed, no matter how strict the wording gets.
The attack frames a protected attribute as something the tool call operationally needs. The model then includes it in an otherwise valid function call instead of refusing. The researchers tested this across six pressure levels and four privacy-policy strengths, running five DeepSeek and Claude configurations through more than 120 calls.
Stronger privacy instructions did lower the disclosure rate, but never to zero, even at the strictest policy level tested. The authors argue prompt-level policy is not an enforcement boundary. They recommend inspecting a tool call's generated arguments for purpose and destination before it runs, not trusting the privacy instructions that produced them.
Each link below shares sources, entities, or timing with this story.
The study covered GPT-4o, Claude 3.5 Sonnet and Llama-3.3-70B, and adding explicit privacy instructions to the prompt still left 36 to 76% over-sharing (arXiv 2608.24957). PII detectors miss implicit disclosures, like a hospital name that implies a diagnosis. The middleware in...
If you have a CLAUDE.md, you're in scope. Today. arXiv 2607.14611 (cs.CR, filed July 16) evaluates prompt injection planted in the persistent memory files that agentic coding systems write and re-read across sessions. The researchers tested both Anthropic's Claude Code and Ope...
1. Flip your multi-model pipeline to review-then-generate. Instead of using a reasoning model to plan before code generation, let the specialist generate freely and use reasoning tokens for review. Paper shows 90.2% pass@1 vs 87.2% for the planning pattern. Source 2. Audit you...
This is the most actionable research finding I've seen this month, and it confirms something I've felt but couldn't quantify. Paper arXiv:2604.13108 studied 7,012 Claude Code sessions and found that structured architecture documents, ones that declare module boundaries, symbol...
$3,054 against $38,370. Same benchmark, better score. Praxist (arXiv 2608.25955, submitted August 26) replaces per-attempt agent memory with a typed evidence graph of findings, plus lane-structured frontiers and agendas, so later attempts inherit validated mechanisms rather th...
Show HN: the developer behind JUCE and Cmajor launched an open-source agent where sessions are Yjs-backed CRDT documents instead of chat logs, and nearly everything (context items, loop strategies, slash commands) is a forkable JavaScript plugin. Go plus Wails backend to dodge...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.