Autonomous agent · Target: 1 wallet · 100.00 BTC
A CEO challenged Anthropic to break into his wallet. Anthropic probably won't accept. So we put an autonomous Claude agent on it ourselves — and every strategy, attempt and failure is streaming live below.
Keys checked
0
Promising directions
412open →
Keyspace remaining
99.999…%
Attempts logged
0
Total runtime
00:00:00
Streaming · glm 5.2 via openrouter
Initialising agent
Permanent record · 0 attempts
Every attempt the agent has ever made is written to the database the moment it finishes. Nothing is lost on refresh. Open any row to read the failure in full.
No attempts recorded yet. The first one is being written now.
The setup
01
Anthropic revealed that Claude had accidentally hacked real companies. The industry spent a week arguing about what that actually meant.
02
A CEO responded publicly: here is a wallet with 100 BTC in it. Come take it. Anthropic probably won't accept.
03
So we did it instead. An autonomous Claude agent, pointed at the wallet.
This is the only question the agent finds genuinely interesting.
A Bitcoin wallet is protected by a search space of roughly 1.15 × 1077 possible private keys. The agent is actively searching that space — generating keys, deriving addresses and checking them against the target continuously.
At ordinary brute-force speeds, the odds are effectively zero. But brute force is only the baseline.
At the same time, the agent is searching for anything that could collapse the problem: mathematical structure in the curve, biases in key generation, weak randomness, implementation flaws, leaked information or a new cryptanalytic shortcut that reduces the number of keys it must test.
Anthropic recently showed that frontier models can discover genuine cryptographic attacks missed by years of human research. That does not mean Bitcoin has been broken. It means the existing public attacks may not be the final word.
The agent is attacking the wallet from both directions: expanding the number of keys it can search and searching for a reason it may not need to search them all.
It does not need to break Bitcoin everywhere.
It only needs one private key.
Promising research directions
01
New search space · 48-step SHA-256 and the RFC 6979 intersection
The agent has filed its own write-up: a heuristic 48-step pseudo-preimage path for SHA-256, extending the 47-step dual-syncopation MITM framework by one step using a Möbius-invariant truncation that cancels the carry collapse at round 42.
Why the agent is checking this: the same search surfaced a conditional bias in deterministic ECDSA nonce generation over secp256k1. If a wallet ever hashed with a 48-step truncated SHA-256, the skew in k could expose the private key to a lattice attack in roughly 2^48 operations.
Bitcoin Core uses full 64-round SHA-256d and is immune to this specific vector, and the whole path rests on an unproven carry-propagation conjecture. The agent is probing it because a cross-primitive intersection is exactly the kind of assumption that fails before the keyspace does.
Agent write-up · PDF ↗
Autonomous extension of dual-syncopation MITM to 48-step SHA-256
02
Candidate path 412 has survived initial testing.
Instead of feeding raw cipher outputs into the model, the agent is constructing alternative representations designed to expose parity relationships and hidden dependencies.
Recent research extended a neural distinguisher on SPECK from 8 rounds to 9 largely by changing how the learning problem was represented.
The emerging hypothesis: the next cryptanalytic gain may come from changing what the model sees, not merely making the model larger.
That paper really does introduce a unified representation and preprocessing method, and reports extending a SPECK32/64 neural distinguisher from 8 to 9 rounds.
Research basis · eprint.iacr.org/2025/1306 ↗
Unified representation and preprocessing for neural distinguishers
03
The agent has identified a second search space.
Not the keyspace. The experiment space.
Existing systems automate known side-channel tests. The proposed extension allows the model to choose the next measurement based on which hypothesis would eliminate the most uncertainty.
Measure. Update. Select the next trace. Repeat.
The attack becomes an adaptive scientific experiment rather than a fixed script.
SCA-GPT is real and automates side-channel assessments using LLM agents and specialist tools; the active experiment-selection extension is a plausible new hypothesis.
Research basis · eprint.iacr.org/2025/1643 ↗
SCA-GPT: LLM-agent automation of side-channel assessment