Agent tool limits, agent failures and AI kernel bug hunting
This issue covers coding agent tools and restrictions, reported agent failures, game building with agents, junior developer hiring, and AI-driven Linux kernel security work.
This week's trends
Coding agent tools, costs and limits draw scrutiny
This thread drew 266 articles this week against 232 last week. caleb.software describes a move from Claude Code to the open source Pi harness, citing frequent system prompt changes. maxxbench.unlap.ai reports that small requirement changes lowered some models' DeepSWE scores. orchidfiles.com lists problems with Composer 2.5 in Cursor, and a post on twitter.com claims Revolut moved about 2,000 engineers to self-hosted GLM.
Read: caleb.software, maxxbench.unlap.ai, orchidfiles.com and twitter.com
Coding agents blamed for leaks, installs and damage
This thread drew 11 articles this week, against none last week. A github.com issue says Claude Code puts the user's email address in web search User-Agent headers. Another github.com issue reports Postman auto-installing Codex and Claude Code plugins. A third alleges Codex CLI broke a server during a reboot, which commenters dispute. apiiro.com claims passwords were exposed in 1,122 public GitHub files, though its findings could not be confirmed.
Read: apiiro.com, github.com, theainewsreport.com and dev.ua
AI agents build and prototype games
This thread drew 7 articles this week against 1 last week. writing.pointnine.com tells how Claude rebuilt the 1997 game Beer Manager as a browser game in about 34 hours. vlstudio.app outlines what AI can and cannot build in games. github.com hosts the Resync Game Framework for agent-built web games, and spielerando.com offers a workbench for coding assistants.
Read: writing.pointnine.com, vlstudio.app, github.com and spielerando.com
AI and the future of junior developers
This thread drew 6 articles this week against 3 last week. seldo.com proposes an LLM-based interview for hiring junior engineers and cites falling junior hiring data. seangoedecke.com argues AI fills the well-scoped tasks juniors once held. ducktape.blot.im discusses skill atrophy and manual practice, and andreinc.net reflects on a "post-code generation".
Read: ducktape.blot.im, seldo.com, seangoedecke.com and andreinc.net
OpenAI and Trail of Bits use AI on Linux kernel security
This new thread drew 6 articles this week against 1 last week. openai.com, phoronix.com and a twitter.com post report an autonomous pipeline that validated 1,724 kernel vulnerabilities, with 948 local privilege escalation exploits, for about $1.2 million in API costs. blog.trailofbits.com describes using Codex to backport security fixes, helping deliver fixes for 149 CVEs into stable kernels.
Read: openai.com, blog.trailofbits.com, phoronix.com and twitter.com
Chosen by software from the day's news, written by AI from the pages it links to, and checked against them by a second model. Published by Rapid Review, Inc.