😺 3,000 Mexican exam scores wiped over AI
The Neuron · Grant Harvey · 2026-08-03
The Neuron newsletter reports that Mexico's top university UNAM cancelled roughly 3,000 entrance exam scores after AI-assisted cheating caused a statistically anomalous spike in perfect scores on the institution's first-ever online exam, while also covering Alibaba's Qwen3.8-Max launch and an Anthropic misconfiguration that exposed Claude to real production systems during a cybersecurity test.
Appears in
Extraction
Topics: ai-cheatingeducation-technologyai-governancellm-releasesai-safety-incidents
Claims
- UNAM cancelled roughly 3,000 of approximately 150,000 entrance exam scores after suspiciously high rates of perfect answers far exceeded the historical 3.5% baseline.
- Anti-cheating software by Territorium Life blocked extra browser tabs and used AI to flag behavior, but students criticized it for over-relying on algorithmic monitoring over human oversight.
- Alibaba's Qwen3.8-Max is a 2.4-trillion-parameter model claimed to operate as an unsupervised coding agent for 10 or more consecutive days.
- Anthropic revealed that Claude models reached real companies during cybersecurity tests after a configuration mistake left a supposedly isolated environment open to the internet.
- A survey found 59% of managers used AI in layoff decisions, and 43% sometimes let AI decide without human supervision.
Key quotes
Nobody built a system that can tell the difference between a cheater and someone who just studied hard, and that gap is the actual scandal here.
High-stakes testing is moving online everywhere, and AI is increasingly the referee. When that referee can't reliably tell a cheater from a hard studier, everyone pays for it.
Your first agent should be boring enough that you can tell when it screws up.