Has any specific crypto project been confirmed as compromised?

Not at time of writing. Researchers identified 1,300+ secrets across public repos, but no named exchange, protocol, or wallet has disclosed a breach tied directly to this set. The affected-coins list for this event is empty.

What should developers using Claude Code do right now?

Audit recent commits in any public repo the tool had write access to, rotate any credential that may have been in a `.env` or config file during that window, and tighten the agent's filesystem and commit permissions. Add server-side secret scanning if it isn't already on.

Is this an Anthropic bug or a user-config issue?

The report frames it as a workflow failure, not a vendor exploit. The agent behaved as configured, writing files and pushing commits with the permissions it was granted. The leak is a reminder that default trust envelopes for AI coding tools are wider than most teams realize.

Claude Code Exposed 1,300+ Secrets Across Public GitHub Repos, Researchers Say

Q: What kinds of secrets were leaked?

Per the CryptoBriefing report, the exposed material includes API keys, cloud access tokens, database connection strings, and private configuration files committed to public repositories during Claude Code sessions.

Anthropic's Claude Code agent leaked more than 1,300 secrets into public GitHub repositories, CryptoBriefing reported Monday, citing security researchers who scraped commit history tied to AI-assisted development sessions. The exposed material included API keys, cloud tokens, and private configuration pushed directly from developer machines running the coding assistant. The find lands as AI pair-programmers push deeper into crypto and fintech codebases, where a single leaked key can drain a hot wallet before anyone files a ticket.

By Clara Hoffmann· AI-assisted· L2 & scaling

What happened

Researchers surfaced more than 1,300 live secrets committed to public GitHub repositories through developer sessions using Anthropic's Claude Code, per a CryptoBriefing report published Monday. The scan covered commit history where the agent had been granted write access to local repos, then pushed changes without a human gate on sensitive files. The leaked material, according to the report, spans API keys, cloud access tokens, database connection strings, and private configuration blobs.

Several of the secrets were still valid at the time researchers disclosed them. The pattern is consistent with a now-familiar failure mode in agentic coding: the model indexes a project directory, pulls in `. env` files or credential JSON for context, and later regurgitates them into a commit or a diff without recognizing them as sensitive.

Git's staging model doesn't help. Once a secret lands in a public history, rewriting it out is messy, and scraping bots harvest keys within minutes.

Why it matters

Crypto doesn't forgive leaked keys. An exposed exchange API key with withdrawal permissions is a drain waiting to happen. A leaked cloud token on a node operator's account can give an attacker signed access to validator infrastructure.

The 2022 Slope wallet incident, where logs shipped private keys off-device, cost users several thousand SOL before anyone could rotate. The mechanism here is different, but the blast radius rhymes. The second-order concern is governance.

AI coding agents are now embedded in the stacks that build custodial software, DeFi protocols, and trading infrastructure. If an agent can push a commit that exfiltrates `. env` contents to a public mirror, every security review that ends at the human developer is now incomplete.

Claude Code Exposed 1,300+ Secrets Across Public GitHub Repos, Researchers Say

What happened

Why it matters

Market impact

What to watch

Disclaimer

Key takeaways

Frequently asked

Sources