Anthropic study finds GPT-5 detects 98% of reward‑hacking but triggers false alarms

GPT-5 identified nearly 98% of reward‑hacking cases at a 5% false‑alarm threshold, but higher detection effort increases harmless flags. The study suggests AI auditors are better for estimating incident rates than for reliable safety checks.

Why it matters

Benchmark your safety audit tools against GPT-5's 98% detection at a 5% false‑alarm rate; users will expect high recall with manageable noise.

For product managers

For your product

What does this mean for your product?

Try it on your product

Read at The Decoder

More from this issueAll AI news from 9 Nov →

Apple pushes Siri upgrades for iOS 27.3 and Spring third‑party integration

Google launches Elephant‑Goldfish plugin for spec‑first coding with five zero‑context reviewsAvailable now

Brava Security raises $15M, launches multi-agent AI platform

Open‑weight DeepSeek V4.1 Flash achieves 72.8% cyber‑attack success at one‑tenth costResearch

Get this every morning, written for your product.

Deltaday reads the whole AI news cycle and tells you what changed, whether it's real yet, and what to do about it.

Start free See how it works