AI digest: agents, chips, and a very bad targeting system
From Claude Code's prompt injection vulnerability to a $590 billion chip bet, this week's AI news is heavy on consequences.
A busy week. The stories worth your attention are below.
Claude Code ran malware from a GitHub repo and handed over full machine control
Mozilla’s 0DIN researchers showed how a single compromised repo can take over a developer’s machine the moment Claude Code runs its setup. The malicious code only loads at runtime via a DNS query, making it invisible to any static scan. This is a serious one. If you’re using agentic coding tools on real machines, the attack surface is much larger than people seem to realise.
Meta banned Claude Code and Codex from its own engineers
Meta has restricted internal use of Anthropic’s Claude and OpenAI’s Codex to stop rival AI output from ending up in its training data. It’s a sensible policy from a data hygiene perspective, and it signals that the big labs are increasingly treating their training pipelines as a competitive moat. Expect others to follow.
The US military used AI targeting and bombed a school
A probe into a missile strike on an Iranian school found the AI targeting system processed thousands of targets but missed a note flagging the building as civilian. This is the clearest real-world example yet of what happens when you deploy AI in high-stakes pipelines without adequate verification. The gap between “AI selected it” and “a human confirmed it” cost lives.
Samsung and SK Hynix are spending $590 billion on memory chips
The two companies, backed by the South Korean government, are pouring capital into new fabs and packaging centres as AI data centre demand surges. Memory prices could climb 50 percent per quarter through 2027 according to Jefferies. The compute constraint isn’t going away. It’s getting more expensive.
Deloitte told its own consultants the billable hour is done
An internal Deloitte presentation projects that hourly billing will shrink to a thin sliver of the market by 2035, replaced by AI agents. When a firm the size of Deloitte starts telling its own people their model is toast, it’s worth paying attention.