DEV Community

Cover image for Agent Memory Myths, Root Escapes, and Rusty Toolchains
Adam
Adam

Posted on

Agent Memory Myths, Root Escapes, and Rusty Toolchains

This week felt like giving our tools a trust fall and then quickly grabbing the safety net. pwning.systems shows how replacing fuzzy agent memory with Datalog-style facts turns "I think this is vulnerable" into auditable reasoning, while Johann Rehberger demonstrates that Claude Code auto mode can still be prompt-injected 60-80% of the time. Add a very real root-escalation bug in Omarchy's defaults and Drydock's npm tarball diffing for supply-chain surprises, and the message is clear: convenience is great, but verify everything.

Performance and infrastructure had a strong week too. Zod 4.5 cut schema memory dramatically, pnpm 12 moved to Rust without breaking team muscle memory, and Postgres WAIT FOR turned stale read-after-write behavior from "almost always" to "effectively never" with low latency impact. Then Sai Srirampur's POSETTE talk reminds us Postgres is often blamed for storage problems it did not create.

On the workflow side, okaneland quantified the MCP context tax, and GitHub Agentic Workflows made the bigger shift explicit: agents are becoming first-class citizens in CI, not just fancy side quests.

Enjoy!

Signup here for the newsletter to get the weekly digest right into your inbox.

Find the 10 highlighted links of weeklyfoo #153:


I accidentally turned LLM memory into program analysis

by pwning.systems

Swapping fuzzy agent memory for a Datalog fact store that supports incremental reasoning in vulnerability research

📰 Good to know, ai,security


Omarchy any user process can escalate to root

by 0xcc

A default Docker configuration let any user process become root - fixed in version 4.0.1

📰 Good to know, security,linux


pnpm 12 rewritten in Rust, same commands and lockfile

by Zoltan Kochan

The Rust rewrite is stable and nearly all v11 commands, flags and lockfile format carry over

📰 Good to know, nodejs,tooling,rust


How Zod 4.5 cut memory per schema by 9x

by Colin McDonnell

Method memoization took a bare z.string() from 7.5KB of heap down to 784 bytes

📰 Good to know, typescript,performance


Breaking Claude Code Opus 5 auto mode

by Johann Rehberger

Prompt injection against auto mode landed 60-80% of the time, so it is a speed bump and not a security boundary

📰 Good to know, ai,security,promptinjection


Read your own writes off the primary with WAIT FOR

by Radim Marek

Naive replica reads were stale 992 times out of 1000 even over loopback - WAIT FOR gets that to zero for 1-2ms

📰 Good to know, postgres,database


The MCP tax - what each server costs before your first prompt

by okaneland

Measuring the context every MCP server eats, and how schema deferral changes the bill

📰 Good to know, ai,mcp,context


GitHub Agentic Workflows

by GitHub

Event-triggered and scheduled AI jobs alongside your deterministic CI/CD, with guardrails and support for Claude Code, Copilot, Gemini and Codex

🧰 Tools, github,ci,ai


Drydock

by Jovi De Croock

Diff built npm tarballs against the last published version, flagging install scripts, network access and new binaries

🧰 Tools, npm,security,supplychain


Postgres is not slow, your storage is

by Sai Srirampur

A POSETTE 2026 talk benchmarking local NVMe against network storage and where the latency compounds

📺 Videos, postgres,performance,video


Want to read more? Check out the full article here.

To sign up for the weekly newsletter, visit weeklyfoo.com.

Top comments (0)