HEAMS benchmarks agent memory for privacy, poisoning and user control
No tested system leads every performance benchmark, and strong task scores can coexist with weak privacy and security.
8 stories, newest first · one card per story
No tested system leads every performance benchmark, and strong task scores can coexist with weak privacy and security.
The design uses a local Horn-clause reasoner and an externally signed knowledge base with k-anonymity or IT-PIR.
One-shot fine-tuning on synthesized episodes with changes and attacks more than doubles evidence-conditioned accuracy against poisoning and outdated facts.
On Trojan Hippo, the authors report zero attack success when self-verification and lightweight verifiers confirm memory.
On MPBench and LoCoMo, attack success was at least 77% below the strongest utility-preserving defense.

Vectorize CEO Chris Latimer says coding-agent memory keeps API keys in plaintext and can be poisoned via plugins and MCP.
MemSentry intercepts persistent-memory writes in agentic AI to catch memory poisoning, reaching 91.7% accuracy with SBERT+LR classification.

Varonis discloses CoSnitch (CVE-2026-24301), three Microsoft Copilot Personal flaws enabling one-click exfiltration of connected-app data; patched August 18, 2026.
Today 2.4K readers · 2.6K views · 142 reads · all time 206K views