An agent’s identity is only as strong as the file it is stored in. On 23 October, Check Point Research published analysis of an infostealer that treats AI agent credentials as a target category of their own.
What happened
Check Point Research documented Aokigahara, a newly observed infostealer (an earlier version may have circulated as Kamidere or Stygian), whose collection routines include a dedicated `aisessions` handler. Alongside the usual browser passwords, cookies, wallets and password-manager data, it goes after authentication files, session data and configuration files belonging to AI tools.
The targeted paths named in the research include:
- Claude Code / Claude Desktop — `~/.claude/credentials.json`, `~/.claude.json`
- Codex — `~/.codex/auth.json`, `config.toml`, `.env`
- ChatGPT Desktop — local app data directories
- Cursor — `state.vscdb` and `storage.json` under global storage
- Windsurf, Kiro, GitHub Copilot / VS Code, Gemini CLI, JetBrains AI Assistant
- Cloud and developer material — AWS credentials and SSO cache, Amazon Q’s local database, Git credential files and SSH private keys
- It also queries the Windows Credential Manager.
Check Point’s assessment of the payload is measured: usable credentials or session tokens “could enable unauthorized access to paid AI services or the resale of compromised accounts,” and the analysis does not confirm how operators monetise what they take. The significance is in the targeting, not in a confirmed campaign against a named victim.
Why it matters for KYA
In most desktop deployments, an agent has no identity of its own — it has a bearer token in a dotfile. Whoever copies that file becomes the agent: same client ID, same scopes, same quota, same audit trail. There is no proof of possession at use time, no attestation that the request came from the machine it was issued to, and nothing in the downstream logs that distinguishes the real agent from a copy running on someone else’s box.
This is the gap between the architecture being built and the machines it runs on. Per-agent SPIFFE identities, X.509 certificates with user-signed delegation assertions, DPoP-bound refresh tokens — all of it assumes private key material cannot simply be lifted by a userland process, because the key stays in a TPM, a secure enclave or a node attestation path. On a developer laptop, the “agent’s key” is a JSON file in the home directory, and every one of those controls degrades to a password.
Stolen agent credentials also break accountability. KYA depends on being able to say which agent acted, under whose authority, in what scope. When a token is shared, copied or resold, the audit trail still looks perfect — it names the agent — while pointing at the wrong principal entirely. The failure mode is not a missing log; it is a confident, wrong one.
The practical response is the same list KYA keeps repeating, now with a threat model attached:
- Shorten lifetimes and bind to the device. Sender-constrained tokens (DPoP, mTLS) turn a stolen file into a stolen file rather than a usable credential.
- Get secrets out of dotfiles. OS keychain, secure enclave, or a brokering service that issues on demand — the agent should hold a handle, not a long-lived secret.
- Treat these paths as secret-bearing. `~/.claude`, `~/.codex`, Cursor global storage and cloud credential directories belong in EDR/DLP coverage and in backup and screenshot hygiene, the same as browser profile data.
- Watch for reuse, not just theft. A token used from a new ASN, country or user agent is the detectable moment; alerting on it is cheaper than any prevention.
- Make revocation per-agent. Rotating one token should not require a password reset, and revoking one agent must not silently disable five others.
- Give each task its own identity. Per-run, least-privilege credentials keep one stolen file to one scope.
None of this is exotic; it is ordinary credential hygiene applied to a new class of principal. The reason it matters is that agent credentials are frequently issued with broader scope and longer life than the human ones they replaced — because the agent is “just automating” something a person used to do.
Caveats: this is malware capability analysis, not a breach disclosure. Check Point describes what the stealer is built to collect and notes that not every targeted file necessarily contains a secret.
Sources
- Check Point Research, “Infostealer riding the AI Train”, 23 October 2026 — research.checkpoint.com.
- Related KYA coverage: Google Cloud’s per-agent SPIFFE identity and secret-free sandbox; AWS’s X.509 agent identity architecture with delegation assertions; Google adding DPoP to OAuth refresh tokens for agents.
Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.