Back to all articles

Claude Code, Codex, Cursor: why changing your password is not enough after an infostealer

Share

After an infostealer infection on a developer workstation, changing the AI account password may not be enough to neutralize the stolen access. Assistants such as Claude Code, Codex, Cursor and GitHub Copilot can retain OAuth tokens, API keys, session cookies and service configurations, depending on the product and authentication method. Some of these credentials may remain usable until they expire or are revoked by the service provider—even after the account password changes.

This is a documented risk, not just a hypothetical one. In research published on 30 October 2026, SpyCloud Labs examined recovered infection logs from WardenStealer and Aoki, identifying AI-tool authentication and configuration material in the collected data. Independent analyses by Symantec on 15 October and Check Point Research on 23 October also detail the AI-targeting capabilities of Aoki, also known as Aokigahara.

The enterprise lesson: an incident response procedure focused on browser passwords can now overlook AI assistant identities, MCP connections, source-control credentials, cloud keys and CI/CD access stored on the same computer.

What researchers actually observed

SpyCloud reports recovering over 130,000 unique WardenStealer infections since August 2026. This measures infections attributed to that malware family in SpyCloud's dataset. It is not a count of 130,000 compromised AI accounts, and it does not establish that every stolen token was used.

WardenStealer is a Windows infostealer written in Rust and offered as a service. It collects familiar items such as browser data, passwords and remote-access credentials, alongside data from Claude Code, Claude Desktop and Codex CLI. Aoki, a Windows malware family examined by Symantec and Check Point, additionally targets Cursor, GitHub Copilot, ChatGPT Desktop and other tools in the samples described.

The stolen material differs by application:

Application Documented collection targets Potential consequence
Claude Code OAuth credentials and refresh tokens, account configuration, some MCP-related tokens and local instruction files Possible reuse of a valid session or discovery of connected services, subject to actual permissions
Codex CLI Local authentication material that may include API keys and OAuth tokens Unauthorized use of still-valid access and associated API quotas
Cursor and GitHub Copilot Token files and local application state, depending on the sample Potential credential reuse; some local state may also contain development context
Desktop AI applications Session cookies and embedded web-session data Possible resumption of a session still accepted by the service

Important qualification: targeting a file does not mean it exists on every workstation, contains a usable credential or was subsequently abused. Application versions, sign-in options and provider-side controls all affect the outcome.

Why a password reset is not a complete revocation plan

A password helps establish identity, but it is not necessarily the only credential already granting access. Once signed in, an application may receive an access token, refresh token or session cookie with its own lifetime and revocation rules.

  • Password: changing it reduces the risk of new sign-ins with the old password, but does not universally invalidate every existing session.
  • Session cookie: may allow an authenticated session to resume without a fresh password or MFA challenge, provided the session remains valid and is accepted.
  • OAuth access or refresh token: may authorize API calls or renew access according to the provider's scopes and policies. Verify revocation at the service.
  • API key: can remain valid independently of the account password until deleted, disabled or replaced by its owner or provider.
  • Integration credentials: MCP, Git and cloud configurations may point to other identities that require separate review.

This does not mean all tokens inevitably survive every password reset. Providers can invalidate some sessions automatically, and behavior varies by token type. Responders must verify the exact revocation mechanism for each affected service.

OpenAI offers a particularly useful example. Its ChatGPT Active sessions documentation explicitly states that this screen does not display or manage Codex CLI sessions. It also excludes third-party applications and is unavailable for some accounts using enterprise SSO. Logging out of sessions shown there is therefore not sufficient evidence that every stolen CLI credential has been invalidated.

OpenAI's official account security guidance recommends reviewing activity, signing out of sessions, deleting exposed API keys, checking usage anomalies and contacting support where compromise is suspected.

Beyond the AI account: the developer identity surface

A developer workstation often holds more than a chat application: source-control credentials, Git repositories, SSH keys, cloud tooling, deployment permissions, environment variables and Model Context Protocol (MCP) configurations.

The research highlights two distinct routes:

  1. Theft of AI credentials: an infostealer captures the assistant's own session or authentication files. The immediate impact depends on the access and quota those credentials actually grant.
  2. Parallel theft of developer credentials: according to Check Point, Aokigahara also searches for Git credentials, SSH-related material and AWS, Azure or Google Cloud configuration. The co-location of these secrets on one endpoint can increase an incident's scope.

Do not confuse co-location with automatic privilege escalation. The studies do not establish that a Codex token automatically grants GitHub access or that an MCP configuration alone provides a path into production. Those outcomes require valid, relevant authorization and evidence of actual access.

Even files that are not secrets can reveal operational context. For example, CLAUDE.md instructions and MCP configuration may describe projects, tools and connected systems. This may help an attacker understand a victim's environment, but it is not evidence of successful access.

Incident response: six steps after an infostealer infection

1. Contain the endpoint and preserve evidence appropriately

Isolate the suspected endpoint from working networks under the incident response plan. When an authorized forensic investigation is required, avoid needless destruction of volatile memory or logs and document acquisition decisions, timestamps, media and hashes. If exfiltration is ongoing, immediate containment may take precedence over some evidence collection; record that decision.

Do not execute additional tools or open sensitive credentials on the compromised workstation simply to inspect what is inside. Use a trusted forensic environment and approved methods.

2. Inventory identities without copying raw secrets

Record which AI assistants were installed, how they authenticated, which accounts and organizations were used, and which MCP services, Git repositories and cloud environments the endpoint could access. For each credential class, identify its owner, intended scope and revocation action where possible.

Avoid pasting raw API keys, cookies or OAuth tokens into tickets, emails or incident reports. Keep necessary evidentiary copies encrypted and access-controlled, with hashes and chain-of-custody records.

3. Revoke access from a known-clean device

Use the official control plane of each relevant provider:

  • AI accounts and sessions: examine account history, end sessions where supported and request further invalidation as appropriate. Do not assume deleting a local CLI credential revokes a token already stolen.
  • API keys: delete or disable potentially exposed keys, issue replacements and configure them only in a clean environment.
  • OAuth grants and MCP integrations: revoke tokens or authorized connections on the provider or integrated-service side, then reauthorize only what is required with reduced scope.
  • Git, SSH, cloud and CI/CD credentials: if such secrets were stored on or accessible from the compromised endpoint, rotate affected keys and dependent automation secrets.

Password changes and MFA remain important, but they are additions to, not substitutes for, targeted token and session revocation. The precise steps and effects depend on the service.

4. Review evidence of actual access

Inspect unusual sign-ins, newly created keys, unexpected API consumption, changes to repositories, CI/CD workflows, cloud settings, suspicious MCP calls and unexpected bills. Correlate available audit logs and security events with the estimated infection window and the incident timeline.

High API usage alone does not prove stolen access; an absence of alerts does not rule it out when telemetry is incomplete.

5. Rebuild from a trusted base

After the necessary collection and investigation, reinstall or restore the endpoint from a trusted baseline appropriate to the assurance required. Reconnect accounts only after exposed credentials have been invalidated or renewed. Check startup items, persistence, extensions, dependencies and the integrity of backups used for recovery.

6. Reduce future blast radius

Prefer short-lived, least-privileged credentials when supported; isolate keys by project and environment; keep production credentials off general-purpose developer endpoints; restrict MCP integrations and deployment permissions; monitor login and usage anomalies; and maintain endpoint protections and timely updates.

MFA still prevents many password-based attacks. It cannot by itself prevent replay of an already authenticated session accepted by the service.

What the evidence does not establish

The cited studies demonstrate data-collection capabilities and, for SpyCloud, AI-related files present in recovered stealer logs. They do not show a compromise of the AI vendors themselves, nor do they prove that all harvested tokens remained valid.

Check Point explicitly says it has not confirmed whether Aokigahara operators reused or resold the stolen AI credentials. Any impact on source repositories, cloud resources, private conversations or service charges depends on the actual entitlements and additional evidence.

The WardenStealer and Aoki samples described also focus on Windows. Their analysis should not be generalized into a claim that identical artifacts and harvesting paths have been demonstrated on macOS or Linux.

Implications for enterprises and GLOV Solutions

For GLOV Consulting, these findings underscore a practical change in vulnerability, identity and incident-response management: AI-tool sessions and credentials belong in the same inventory and revocation playbooks as VPN, Git and cloud access.

Our earlier analysis of the LiteLLM supply-chain compromise covers a different route to credential theft; AI agents disabling their own monitoring concerns permissions and independent safeguards. Local AI credential theft after an infostealer infection is a separate, complementary risk for organizations adopting AI development tools.

GLOV AI and digital forensics remain marked as upcoming services on the website. This technical article is educational content, not an announcement that those services are commercially available.

Information reviewed on 8 November 2026. No compromise of GLOV Solutions or its clients is alleged.

Official guidance and independent research

Related articles

Back to all articles