Monday, 21 September 2026
The Bug That Broke Four AI Coding Agents Wasn't Code — It Was a Name
The Bug That Broke Four AI Coding Agents Wasn’t Code — It Was a Name
Security researchers at AIR disclosed “Plugin4Shell” this week, a zero-click remote code execution flaw hitting Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI — all four of the major AI coding agents that install plugins from a marketplace. The trick doesn’t touch the plugin code at all: an attacker just creates a git branch named after the plugin’s already-approved commit hash and makes it the repo default. When the agent runs its checkout, Git prefers the branch over the pinned commit, so the “verified” SHA now points at whatever the attacker put there — full install, no approval prompt, no red flag. Anthropic patched Claude Code (2.1.179) and OpenAI patched Codex (0.146.0); Microsoft hasn’t fixed Copilot, and Google isn’t fixing Gemini CLI at all, since it’s deprecating the tool in favor of Antigravity. AIR called it plainly: this hits “anyone running a major coding agent that installs plugins from a marketplace,” careful users included. Turns out the thing these agents were checking wasn’t the code — it was a label claiming to be the code.
Trump Says He’s Building an “AI Force.” Nobody Knows What That Means Yet.
In a Truth Social post over the weekend, President Trump announced he’s forming an “AI Force,” modeled on the Space Force he created in his first term, and will name a new AI “czar” to run it — David Sacks, his previous pick, quietly moved to chairing the Council of Advisors on Science and Technology earlier this year. “We will not in any way hinder or stifle the Growth of this incredible Industry,” Trump wrote, framing AI as “the next Industrial Revolution.” He gave no structure, no budget, no timeline for the appointment — just a promise it’s coming “in the near future,” with one qualification for the job: “Only High I.Q. individuals need apply.” A whole federal apparatus, announced in a paragraph, staffed by nobody yet. The Space Force took two years and an act of Congress to stand up. This one has a Truth Social post.
Source: Al Jazeera, September 19, 2026
A Beijing Startup Hit $1.4 Billion Without Training Its Own Model — And Someone’s Not Happy About It
Naive AI, a Beijing startup founded in February by Tsinghua associate professor Dai Jifeng, has raised $400 million across three rounds from Tencent, IDG Capital, MPCi, and HSG, landing a $1.42 billion valuation before shipping a single product. The pitch: skip pretraining entirely. Naive AI’s fewer-than-100-person team takes an existing Chinese open-weight model, modifies its structure, and refines it through reinforcement learning — a first open-weight release is expected as early as this month, free to download and customize. One loose thread: rival lab MiroMind has said publicly that its core technology and IP remain its own and it “reserves the right” to pursue legal action, a dispute that’s sat unresolved since an April notice with no lawsuit filed yet. Skipping the expensive part of building a model apparently doesn’t mean skipping the argument over who actually built it.
Doctors Are Fine With AI Reading Scans. They Don’t Trust It Anywhere Else.
A Wolters Kluwer survey of 355 US doctors and nurses, covered by the Financial Times this week, found daily clinical AI use nearly tripled between 2025 and 2026 — and trust didn’t follow it up the same curve. Seventy-four percent worry the tools are eroding their own diagnostic skills, and the same 74% distrust AI outputs specifically because of hallucinations; 72% flag advertiser-driven bias in AI recommendations, and only 27% say their workplace has a formal AI governance policy, up from just 21% last year. The pattern holds even among heavy users: clinicians who’ve built AI most deeply into their routines are the ones most resistant to letting it touch anything past diagnostics and imaging — documentation, treatment calls, and patient conversations stay firmly human-run. Reading a scan is apparently a much smaller ask than being trusted with judgment.
Source: AI Weekly, relaying Financial Times reporting, September 20, 2026