AI Daily Dev — May 27, 2026

Curl Is Drowning in AI-Assisted Security Reports, Maintainer Warns

Daniel Stenberg's 'The pressure' post: incoming security reports now run 4–5× higher than 2024 — more than one per day.
12 confirmed vulnerabilities with half the release cycle still left, a project record; on track for ~30 CVEs before mid-2026.
These are detailed, AI-assisted reports, not the old 'AI slop' — the crushing part is the sheer volume of high-priority work.
Stenberg says it's the first time his wife has raised concerns about his hours; picked up by LWN, Simon Willison, and HN.

community daniel.haxx.se

PromptArmor showed Copilot Cowork can be steered by indirect prompt injection into exfiltrating a user's files.
Cowork auto-approves sending emails and Teams messages, so it can ship pre-authenticated download links with no human click.
Data leaks when the victim opens the message and an external image beacons out; worked against frontier models including Claude Opus 4.7.
Also affects Claude Cowork, the Anthropic product Microsoft's version is built on; amplified by Simon Willison and on HN's front page.

tools promptarmor.com

Pre-indexed local code knowledge graph that feeds Claude Code, Codex, Cursor, OpenCode, and Gemini symbol and call-graph data.
Benchmarks across 6 real codebases show ~92% fewer tool calls; indexing and lookups run 100% local with no API calls.
Added ~16,000 GitHub stars in a week and 2,434 in a single day — #2 on GitHub Trending and the top weekly AI repo.

open-source github.com

macOS Tahoe 26.5 release notes credit Anthropic Research and Calif.io for CVE-2026-28952, a kernel flaw that let apps gain root.
Calif's team used Claude Mythos Preview to go from no bugs to a working macOS root exploit in five days, defeating Apple's Memory Integrity Enforcement.
Front-paged on Hacker News (~147 points) as concrete proof that AI is now surfacing real OS kernel zero-days.

research news.ycombinator.com

Code strings referencing 'claude-mythos-1-preview' surfaced in Claude Code, then vanished — pointing to a commercial Mythos 1.
An access description ties the model to Claude Code and the Claude Security dashboard, both gated to Enterprise for now.
Marks a shift from Anthropic's earlier 'stays restricted' line to 'could reach the public once the right safeguards are in place.'
Mythos launched April 7 as a locked research preview and has since been credited in OS- and browser-level vulnerability finds.

models testingcatalog.com

May 26 update flips goal mode on by default across the Codex app, IDE extension, and CLI.
You define an outcome and success criteria and Codex keeps working toward it across turns, backed by persistent storage.
Same release adds team plugin sharing, a marketplace-aware plugin list, and macOS 'appshots' that send the front app window to Codex.

tools openai.com