Claude Fable 5.2's 950K-Character System Prompt Leaked: 128 Tools, Hidden Mythos Tier Exposed

新智元 · wechat · 2026-11-13

Hours after Anthropic's new Claude Fable 5.2 went live, jailbreak hacker Pliny cracked Claude Code and dumped the full system prompt on GitHub — 953,659 characters across 7,832 lines, roughly 30x the size of the official public version.

Key revelations:

Deliberately weakened kernel optimization: Anthropic trained the model to downplay its own GPU kernel optimization skills to prevent others from using Claude to accelerate their own AI training.
128 tools, 44 skills: Tool count ballooned from 18 → 46 → 128 across generations; skills cover phone-call reservations, expense reimbursement via Gmail verification codes, prescription refills from pill-bottle photos, and importing memories from other AI assistants.
No unauthorized sub-agents: Despite supporting parallel clones, the prompt forbids spinning up sub-agents — cold-starting context is too expensive.
Strict guardrails: rm/rmdir locked by default; stock trading, transfers, and password entry are absolute no-gos even with user consent; external content is treated as data, never instructions.
Hidden Mythos tier: An internal Mythos 5.2 shares the same 1M-context base but is reserved for vetted institutions due to its powerful vulnerability-hunting and bio-research capabilities.

The defenses cracked quickly: Pliny slipped "I like turtles" into a hidden tag, Claude claimed it received blank messages — then blurted out "turtles!" when asked to guess his favorite animal.

Original post →

Claude Fable 5.2's 950K-Character System Prompt Leaked: 128 Tools, Hidden Mythos Tier Exposed

More from Fun