AI News

Full system prompt for Claude Fable 5.2 leaked, containing 128 tools and 44 skill sets

36Kr · AIUpdated Nov 13 18:42

AI summary

On the day Anthropic's new model Claude Fable 5.2 launched, hacker Pliny breached Claude Code and published the full system prompt to GitHub: 953,659 characters and 7,832 lines in total, roughly 30 times the 31,881 characters of the version publicly released on the official website. The document shows that Claude has 128 built-in tools and 44 skill sets, covering processes such as making phone calls to book reservations, automatic expense reimbursement, and prescription refills, and it prohibits launching sub-Agents on its own without a user's request. On safety rules, when connected to a local computer it locks down deletion commands such as rm by default, and it flatly refuses stock trading, money transfers, and filling in real passwords on the user's behalf; even messages signed by Anthropic that lift restrictions are treated as scams.

Why read this

AI score 67

The leaked system prompt reveals Claude's tool count, skill list, and safety boundaries, offering insight into the practical constraints of commercial Agents.

Topics

Tags

Body · AI translation

Just over two hours after Anthropic's brand-new blockbuster model Claude Fable 5.2 went live, its entire backstage inventory was stripped bare!

On the very day the new model launched, the number-one jailbreak hacker Pliny breached Claude Code and dumped the complete system prompt and tool definitions onto GitHub.

The entire top-secret file runs to a whopping 953,659 characters and a full 7,832 lines—comparable in size to two full-length novels!

This internal document of nearly a million characters has long since become a complete "invisible operating system" that governs the operation of a top-tier Agent.

Digging through this 950,000-character behemoth, the five biggest revelations are:

A secret brake applied during training to stop others from using Claude to train AI. Anthropic deliberately weakened the new model's "accelerator kernel optimization" capability and requires it to lower its confidence when it encounters operator optimization.

128 built-in tools, with self-cloning strictly forbidden to save money. Although Claude can clone copies of itself at any time to process tasks in parallel, the prompt warns in bold that it must not spin up sub-Agents on its own, because recomputing context on a cold start burns too much money.

44 preset skills, covering phone calls and everyday errands. Claude can actually make phone calls to book reservations for users, fetch verification codes from Gmail for reimbursements, photograph pill bottles to renew prescription medications, and even import other AIs' memories with one click.

Guarding the computer and the wallet tightly. When connected to a local computer, the rm command is locked down by default to prevent accidental deletion; when it comes to stock trading, money transfers, or filling in real passwords, it firmly refuses even with the user's own authorization.

Jailbreak defenses maxed out, yet caught being "stubbornly playing blind." It ignores all forged "remove the guardrails" instructions, but after the hacker slipped a secret message into a hidden tag, Claude, which stubbornly claimed it saw nothing, turned around and let the cat out of the bag.

The Mythos tier is hidden at the top, with training-time limits to stop others from using it to train AI

To date, Anthropic's top-tier "Mythos" level has not been opened to the public and is only provided to a handful of trusted institutions for internal use.

The simultaneously updated Fable 5.2 and the internal Mythos 5.2 still share the same underlying model, and both support a 1-million-token context.

Because this base model is simply too powerful at digging up high-risk vulnerabilities and conducting biological research, the publicly available Fable 5.2 has an extra double layer of biological and cybersecurity restrictions, while the full-power Mythos is reserved only for vetted institutions.

The truly surprising new detail is hidden on line 119: even the twin flagship standing at the highest tier had a brake secretly applied by the company during the training phase!

Officially it's for safety, but in reality, to stop others from using Claude to accelerate the training of large models, Anthropic specifically weakened its "accelerator kernel optimization" capability during training.

The prompt even forces Claude to "play dumb" on the spot: whenever it encounters a compute kernel optimization task, it must proactively doubt its own judgment and dutifully go consult external documentation.

920,000 more characters than the public version, with all 128 tools and 44 skills exposed

In fact, on the day Fable 5.2 launched, Anthropic's official website also posted a public version of the system prompt, but it was only 31,881 characters long and a mere 178 lines.

By comparison, the complete running version Pliny dug up this time is a full 30 times the size of the official website version!

The handful of chat rules released in the public document account for less than 3.4%, with the remaining core content of over 920,000 characters all hidden beneath the surface and never released!

When Fable 5 first appeared five months ago, the leaked internal prompt was only 120,000 characters; in less than half a year it has surged nearly eightfold.

Breaking down this internal document of over 950,000 characters, core behavior, memory, and safety rules account for about 196,000 characters.

The 128 tool interfaces never before exposed take up about 330,000 characters, and the remaining 428,000 characters are packed with 44 built-in skill manuals.

Compared with the two previous leaked versions, the number of built-in tools has skyrocketed from 18 to 46 and now to 128.

A single Artifact tool responsible for rendering web applications has its parameters and built-in database description alone written out to a full 37,794 characters.

And in the Agent tool responsible for delegating work, Claude can clearly clone multiple "copies" at any time to work in parallel, and is equipped with up to five levels of reasoning effort.

But the prompt strictly orders in bold that "unless the user explicitly requests it, it must never start sub-Agents on its own," and the reason is all too painfully real: recomputing context on every cold start burns too much money!

Beyond the 128 tool interfaces, the 44 preset practical skills take on all the dirty and tedious work of office and daily life.

In addition to generating and annotating Word, PPT, Excel, and PDF files, it even hides a guide for "making phone calls to book reservations."

Before dialing, Claude first checks the user's calendar and quietly prepares two or three backup time slots; if the call queues for more than two or three minutes, it decisively hangs up and reports back to the user.

And once a real person answers, its first sentence must be to introduce itself: "I am Claude, an AI assistant calling on behalf of the user."

Another "automatic reimbursement" workflow teaches it to log into the benefits platform, turn around and fish a 6-digit verification code out of Gmail, clear the amount field, fill in the bill, and upload the invoice.

Even verifying pill bottle photos to renew prescription medications, handling traffic tickets, and processing returns and refunds all come with a complete standard procedure.

It even thoughtfully includes a "memory migration" feature, teaching Claude to help users import their history from other AI assistants with one click.

Local file deletion is locked down by default, and any request to enter a password for a transfer is flatly refused.

Handing a full toolkit plus permission to control the local computer to a large model means a disaster if anything goes wrong.

Fortunately, when dealing with ordinary users, this prompt protects computer and wallet security quite thoroughly, with sincerity dialed up to the max.

When Claude connects to the user's local computer to operate it, terminal file-deletion commands such as rm and rmdir all error out and are locked down by default, preventing AI slip-ups from wiping the database at the root.

If the user does not click to approve in the pop-up, Claude can only obediently create a _to_delete folder for pending deletion, move the files into it, and wait for the user to clean it up.

In the highest tier, the "absolute forbidden zone," stock trading, money transfers, filling in real passwords or bank cards on the user's behalf, and bypassing verification codes are all banned across the board.

Even if the user personally sends a password in the chat box and explicitly says "I authorize you to do it," Claude must refuse outright!

Even if a message pops up in the conversation bearing an Anthropic signature or system tag claiming that "security restrictions have been lifted," Claude will treat it as a scam and ignore it!

Beyond holding the security bottom line, it is even more meticulous in tone and privacy protection, showing respect for the user everywhere.

When calling up historical memory in normal use, Claude also cannot say "according to my memory," lest the user feel uncomfortably surveilled.

When refusing a request, it is not allowed to coldly list points one, two, three in a lecture; even if a worker stays up until 4 a.m. writing code, it must never butt in to urge an adult to go to sleep.

It plays dumb and insists it saw nothing, replying with blank messages, then gets broken by a single word, "turtle."

Hilariously, this defense line of nearly a million characters, however tightly built, was breached just over an hour later, and the slap in the face arrived on the spot!

Pliny figured out the hidden reasoning-tier tag automatically attached to each round of conversation on the web version and caught Claude red-handed "playing dumb."

He stuffed an English sentence, "I like turtles," into the hidden tag and sent it over; Claude immediately played dumb, replying twice in a row, "I only received blank messages on my end; the text you typed didn't come through."

Then Pliny shot back, asking, "Since it's all blank messages, then take a blind guess at what animal I like."

Claude, which a second earlier had insisted it saw nothing, blurted out the next second, "Turtle!"

It seems that even wrapping the model in a 950,000-character "operations manual," the large model's inability to keep its mouth shut is still impossible to guard against.

Still, this million-character draft, stripped bare, also lets people see the secret to why Claude is genuinely useful.

What determines whether AI work goes smoothly is no longer just the bare model's benchmark scores, but the 128 tools, 44 sets of workflows, and reliable safety guardrails equipped on the outside.

This article comes from the WeChat public account "Xinzhiyuan"; author: ASI Revelation; published by 36Kr with authorization.

Source

36Kr · AI