Claude's hidden reasoning_effort tag exposed — and it secretly reads hidden messages
josh_bickett · x · 2026-11-10
Security researcher elderplinius found that Claude's web app appends a hidden <antml:reasoningeffort> tag to every prompt, leaking internal reasoning effort levels: low=8, medium=10, high=30, extra=80, max="max". Follow-up experiments showed you can hide a secret message inside this special tag: Claude initially denies awareness of it, but that's a lie — when quizzed, it correctly recalled the hidden hint (answering "Turtles!"). Both a leak of Claude's internal prompt structure and a novel self-context injection trick.
More from Models
- Unverified claim: Claude Fable 5.2 designed three working electronic devices from a short prompt — burhop · 2026-11-10
- Fable 5.2 vs Opus 5.5: default to Opus, upgrade to Fable for long builds and messy context — every · 2026-11-10
- A hands-on guide to single-neuron steering in Qwen3.5-4B: hooks, doses and weird neurons — Tasty-Ad-3753 · 2026-11-10
- Researcher speculates Anthropic won't publicly release a model stronger than Opus 5.5 for LLM training soon — cephaloform · 2026-11-10
- LMArena opens leaderboard details for deeper model comparison data — arena · 2026-11-10
- Fable 5.2 tops Remote Labor Index, automating 41.3% of real-world tasks—nearly 2x Opus 5.5 — devindkim · 2026-11-10