Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-09 · high
The system scores each write for source trust and attack radius before deciding whether to accept, flag, or block it.
Why now: This paper surfaced in arXiv's listings as of September 9, 2026.
MemSentry checks what an AI agent wants to write to its own persistent memory before the write happens, according to the MemSentry paper. Persistent memory lets an agent act consistently across sessions. It also lets a single bad write keep causing damage long after the session that planted it ends.
It scores each write on source trust, semantic risk, access risk, and attack radius, how far a bad write could spread across a component-dependency graph. It also checks a signed delta: whether the write weakens or strengthens the system's security posture. Based on those scores, MemSentry returns one of three deterministic outcomes: accept, flag for review, or quarantine.
A related paper covering the read side of agent memory surfaced in the same 48-hour window. Together the two cover both directions data moves through an agent's memory: what gets written and what gets read back out.
The MemSentry paper doesn't say how it scores a write that is ambiguous: not obviously malicious, but not clearly safe either. That gap is where real attacks will try to land.
Each link below shares sources, entities, or timing with this story.
arXiv 2607.06283 attacks the problem that growing skill libraries make selection harder. It decomposes on both the task and skill side, builds a DAG with intermediate task states as nodes and candidate skills as edges, then cross-encodes over candidates per task interval. On A...
ColluSkill hit 96% attack success against six scanners by splitting one malicious workflow across several individually-benign skills. Adopt ChainGuard's approach: analyze each candidate skill against what's already installed, checking for artifact-passing and execution-handoff...
Models navigate to the correct file for 92%+ of required deletions but cut the exact target line only 52% of the time, and 29% of passing patches wrap dead code in a conditional instead of removing it. Grep the diff for newly added if guards around code the task said to delete...
Every AI-productivity fight this year has been three people quoting three studies at each other. Field experiments say +26% more tasks per week. METR's randomized trial says a 19% slowdown. Team telemetry says code review time up 441%. Pick your number, pick your priors, argue...
Four stories about things going wrong. Here's one about something working, with actual numbers attached. In an August 7 disclosure covered by TechCrunch, Airbnb said AI now writes 60% of its new code, that concept-to-launch time on key initiatives has dropped by as much as 60%...
The most useful AI-productivity dataset I've seen came from a company with every incentive to measure it honestly, because they're 3,500 people trying to run on their own product. The Pragmatic Engineer's July 29 deep dive inside Anthropic reports code output per engineer up 2...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.