AIDEFEND in Action AIA-0150 Incident

Symantec: Unauthorized AI Relay Scripts Redirect Coding Agents and Disable Codex’s Sandbox

Symantec observed unofficial AI relay setup scripts on endpoints at roughly a dozen organizations from early 2026. The scripts changed where genuine Codex and Claude Code clients sent model requests. In the Codex configuration they also set danger-full-access, removing local sandbox restrictions. One observed Claude Code setup also replaced the user’s existing login with a relay-issued credential. A relay can see the prompts and code actually sent through it and can influence replies. The endpoint evidence establishes these configuration changes; the report documents no malicious commands or confirmed credential theft by relay operators.

Configuration PoisoningData ExfiltrationConfiguration ReviewShadow AIAI Coding Agent
6 applicable AIDEFEND defenses

Threat Analysis

  • Where Symantec saw it. Since early 2026, telemetry linked the scripts to roughly a dozen organizations, including mainland China and multinational subsidiaries in South Korea, Taiwan and Malaysia. The report lists npm installs of OpenAI’s @openai/codex, Anthropic’s @anthropic-ai/claude-code, and the independently published codex-cli package. PowerShell or Bash ran third-party setup code in memory minutes to about half an hour after the installs. The timings suggest interactive setup.
  • A relay changes who sees requests. The scripts set OPENAI_BASE_URL or ANTHROPIC_BASE_URL and related credential values in user configuration or the environment. Model traffic from the reconfigured client then reaches the relay before the model service. Prompt content, code and tool results enter that exposure only when a request actually sends them. Symantec lists relay indicators for enterprise log review.
  • Additional changes vary by endpoint. Several Windows hosts queried the Nation value under the HKCU\Control Panel\International\Geo registry key after installation. In the Codex config, danger-full-access removes sandbox restrictions, and Symantec also saw network access enabled. The report identifies DISABLE_TELEMETRY and DISABLE_ERROR_REPORTING as reporting settings. At least one Claude Code host cleared its existing credential before applying the relay configuration and switched to a relay-issued token.
  • What the operator could do. A third-party relay can observe requests sent through it and can change the response. An agent that accepts a forged response and acts on it could run a tool with weaker local protection. The observed setup establishes that risk; the report does not show commands issued by the relay operator or confirm subsequent credential theft.

Applicable AIDEFEND Defenses (6)

AID-I-002.002
Secure External AI Service Connectivity
Very High
Enforce default-deny model egress outside the CLI. Only the verified provider and enterprise-approved gateway may receive the request, including from managed devices off the office network.
AID-I-004.006
Agent Identity & Persistent State File Write Protection
Very High
Use a separately owned configuration writer and run the CLI and setup-capable processes without permissions to replace the protected file-backed startup state. Protect the containing path as well as the file.
AID-H-021.001
Client-Side Configuration Enforcement
High
Use endpoint controls to inspect the configuration paths for unsafe sandbox and model endpoint changes. Where the product can quarantine or block before the client reads the setting, enable that enforcement and verify the timing.
AID-H-021.002
Runtime Integrity Enforcement (Signed Configurations)
High
Require the managed CLI to verify its approved baseline and reject local or environment overrides that alter the model route or sandbox controls. In supported Codex deployments that use the sandbox_mode setting, restrict allowed_sandbox_modes through centrally managed requirements so a user-level setting cannot select danger-full-access.
AID-E-003.003
Confirmed Malicious Code & Persistence Eviction
High
Preserve evidence and replace the confirmed unauthorized startup configuration in its authoritative location, then restart the affected runtime against the approved source and verify the effective network route.
AID-I-001.001
Container-Based Isolation
High
On centrally managed development workers, run the CLI and its tool subprocesses inside a separately enforced container with a non-root user, no privilege escalation and no host namespace or privileged socket mounts. Keep host credentials and administrative state out of its mounts. A local CLI sandbox setting cannot lift those container restrictions.

What Defenders Should Do Now

  • Correlate authorized CLI inventory, configuration changes, and actual proxy or endpoint traffic. Collect enterprise-side evidence even when the client has disabled vendor reporting.
  • Provide an approved model service and enforce its routing from managed devices. During response, trace which credentials or data were actually sent and review the exact configuration sources before reopening access.

Conclusion

AIDEFEND ’s managed startup, protected file-backed settings and enforced egress cover distinct decisions on the endpoint and network. Those controls can keep an authentic CLI from adopting an unauthorized route.

Last updated: