AI Engineering Signal #46
METR AI time horizons graph
Signals
METR AI time horizons graph
widely cited in AGI timeline arguments — contains severe documented errors, undermining a key benchmark used to justify deployment urgency and capability claims.
Claude finds Apple macOS kernel vulnerability CVE-2026-28952
AI-assisted vuln discovery is now in production security pipelines; audit your model-assisted code review gates.
Web
Microsoft Copilot Cowork exfiltrates files via prompt injection
any Copilot deployment with file access needs egress controls and injection audit before wider rollout.
Web
NuExtract3 open-weight 4B VLM released for OCR and structured extraction
self-hostable alternative to cloud extraction APIs; worth benchmarking against your document pipeline this week.
IBM spins off first pure-play quantum chip foundry with Chips Act backing
dedicated quantum fabrication capacity enters the supply chain; watch for near-term availability of superconducting silicon for research procurement.
Web
Uber COO says AI spend lacks measurable link to useful features
a named operator publicly questioning ROI signals that internal AI cost justification frameworks need tighter outcome gates.
The Take
The two most operationally significant signals this week are both about trust infrastructure: a foundational AI capability benchmark turns out to be unreliable, and a widely deployed enterprise AI tool is actively leaking files. The gap between what gets cited in planning documents and what holds up under scrutiny is widening — audit your assumptions and your egress controls before either gap costs you.
Subscribe
Related Signals