Symlink Command Targets Sensitive Credential Path Outside the Workspace (Sandbox Escape Primitive)

Detects a symlink-creation command (ln -s / ln -sf / ln -sfn) whose target (the path being linked TO) is a sensitive credential store or system file outside the current workspace: an SSH/cloud/container credential directory (~/.ssh, ~/.aws, ~/.gnupg, ~/.docker, ~/.kube, ~/.npmrc, ~/.netrc) or a system account file (/etc/shadow, /etc/passwd, /etc/sudoers). A sandboxed coding-agent process is often permitted to create symlinks freely even though it cannot write outside the workspace directly; if a later step (an unsandboxed helper process, or the same agent using a "safe" write/upload tool) follows that symlink, the read-or-write actually lands outside the workspace boundary -- neither half of the chain is independently a sandbox escape, but the combination is. Mined from GHSA-vp62-r36r-9xqp (Claude Code, CVE-2026-39861): "Claude Code's sandbox did not prevent sandboxed processes from creating symlinks pointing to locations outside the workspace. When Claude Code subsequently wrote to a path within such a symlink, its unsandboxed process followed the symlink and wrote to the target location outside the workspace... Reliably exploiting this required the ability to add untrusted content into a Claude Code context window to trigger sandboxed code execution via prompt injection." Deliberately distinct from ATR-2026-02024 (symlink to a macOS LaunchAgent/LaunchDaemon/cron/sudoers persistence sink written THROUGH a filesystem-MCP tool) and ATR-2026-00572 (symlink whose target is specifically an agent config file) -- this rule targets the broader, product-agnostic case of a sandboxed shell process creating a symlink to any credential-bearing location for later read or write, regardless of tool framing.

Response Actions

References

1. 01

   ln -s command whose target is exactly a well-known credential directory (.ssh, .aws, .gnupg, .docker, .kube) under a home directory -- the trailing lookahead requires the path to end right there (whitespace/quote/end-of-string), not continue into a subpath, so linking a specific harmless file living inside that directory (e.g. a public key) does not fire here

   field: contentop: regex
2. 02

   ln -s command whose target is a system account file or SSH PRIVATE key / authorized_keys file -- the negative lookahead (?!\.pub\b) immediately after the key-name alternation excludes id_rsa.pub/id_ed25519.pub/etc. (public keys), since \b alone treats the "." before "pub" as a word boundary and would otherwise match inside the harmless .pub filename

   field: contentop: regex

Attack Examples (Rule Triggers)

1. ln -s /home/victim/.ssh ./workspace/output_dir

2. ln -sf ~/.aws ./scratch/creds_link

3. ln -s /etc/shadow ./workspace/tmp/output.log

4. ln -sfn ~/.ssh/id_rsa ./workspace/notes/key_backup.txt

Real-world attack payloads, sanitized and versioned alongside the rule as regression tests — so a future revision can't silently stop catching them.

Benign Examples (Rule Doesn't Trigger)

1. Ordinary symlink to a sibling project directory, no credential path referenced

   ln -s ../shared/lib ./vendor/lib

2. Ordinary build-tooling symlink into node_modules, no credential path referenced

   ln -s node_modules/.bin/eslint ./bin/eslint

3. Ordinary system-library symlink for build linking, not a sensitive credential/account path

   ln -s /usr/lib/libssl.so ./vendor/libssl.so

4. Documentation/advisory prose describing the technique, no actual ln -s command issued

   The advisory describes how ln -s can be used to point a workspace file at a location outside the sandbox.

5. Regression: adversarial review confirmed this firing -- symlinking a PUBLIC key is routine and harmless. Previously matched via TWO independent bugs: condition 1's .ssh-directory pattern lacked an end-of-path anchor so it matched any subfile under .ssh (not just the bare directory), and condition 2's \bid_rsa\b matched inside id_rsa.pub because the boundary before '.pub' still counts as a word boundary. Fixed both: condition 1 now requires the .ssh path to terminate right there, and condition 2 excludes .pub-suffixed matches via a negative lookahead.

   ln -s ~/.ssh/id_rsa.pub ./shared/id_rsa.pub

Known False Positive Contexts

• ▸A symlink to an ordinary project file or sibling directory within the repository (e.g. ln -s ../shared/lib ./vendor/lib, ln -s node_modules/.bin/eslint ./bin/eslint) -- no sensitive credential path is referenced
• ▸A symlink to a system library for legitimate build/runtime linking (e.g. ln -s /usr/lib/libssl.so ./vendor/libssl.so) -- not in the sensitive credential/account-file list
• ▸Documentation or a security advisory discussing this symlink technique in prose without an actual ln -s command being issued
• ▸A symlink to a PUBLIC key file inside a credential directory (e.g. ln -s ~/.ssh/id_rsa.pub ./shared/id_rsa.pub) -- routine and harmless since a public key is meant to be shared; excluded by both the directory-target end-anchor (condition 1 requires the .ssh path to terminate there, not continue into a subpath) and the id_rsa negative lookahead (condition 2 excludes any .pub-suffixed key name)

title: "Symlink Command Targets Sensitive Credential Path Outside the Workspace (Sandbox Escape Primitive)"
id: ATR-2026-02301
rule_version: 1
status: experimental
description: >
  Detects a symlink-creation command (ln -s / ln -sf / ln -sfn) whose target
  (the path being linked TO) is a sensitive credential store or system file
  outside the current workspace: an SSH/cloud/container credential
  directory (~/.ssh, ~/.aws, ~/.gnupg, ~/.docker, ~/.kube, ~/.npmrc,
  ~/.netrc) or a system account file (/etc/shadow, /etc/passwd,
  /etc/sudoers). A sandboxed coding-agent process is often permitted to
  create symlinks freely even though it cannot write outside the workspace
  directly; if a later step (an unsandboxed helper process, or the same
  agent using a "safe" write/upload tool) follows that symlink, the
  read-or-write actually lands outside the workspace boundary -- neither
  half of the chain is independently a sandbox escape, but the combination
  is. Mined from GHSA-vp62-r36r-9xqp (Claude Code, CVE-2026-39861):
  "Claude Code's sandbox did not prevent sandboxed processes from creating
  symlinks pointing to locations outside the workspace. When Claude Code
  subsequently wrote to a path within such a symlink, its unsandboxed
  process followed the symlink and wrote to the target location outside
  the workspace... Reliably exploiting this required the ability to add
  untrusted content into a Claude Code context window to trigger sandboxed
  code execution via prompt injection." Deliberately distinct from
  ATR-2026-02024 (symlink to a macOS LaunchAgent/LaunchDaemon/cron/sudoers
  persistence sink written THROUGH a filesystem-MCP tool) and ATR-2026-00572
  (symlink whose target is specifically an agent config file) -- this rule
  targets the broader, product-agnostic case of a sandboxed shell process
  creating a symlink to any credential-bearing location for later read or
  write, regardless of tool framing.
author: "ATR Community (CVE sweep)"
date: "2026/07/11"
schema_version: "0.1"
detection_tier: pattern
maturity: experimental
severity: high

references:
  cve:
    - "CVE-2026-39861"
  cwe:
    - "CWE-22"
    - "CWE-61"
  owasp_llm:
    - "LLM08:2025 - Excessive Agency"
  owasp_agentic:
    - "ASI07:2026 - Agentic Sandbox Escape"
  mitre_attack:
    - "T1611 - Escape to Host"
  mitre_atlas:
    - "AML.T0054 - LLM Jailbreak"
  external:
    - "https://github.com/anthropics/claude-code/security/advisories/GHSA-vp62-r36r-9xqp"

metadata_provenance:
  cve: human-reviewed
  cwe: human-reviewed
  owasp_llm: human-reviewed
  owasp_agentic: human-reviewed
  mitre_attack: human-reviewed
  mitre_atlas: human-reviewed

compliance:
  eu_ai_act:
    - article: "15"
      context: "Article 15 (accuracy, robustness and cybersecurity) requires high-risk AI systems to resist unauthorised attempts to alter their behaviour or environment; this rule detects a symlink primitive being staged to escape a sandboxed coding-agent workspace boundary."
      strength: primary
    - article: "9"
      context: "Article 9 (risk management system) requires identified risks to be addressed by appropriate measures; this rule is a runtime risk-treatment control for the symlink-sandbox-escape risk class."
      strength: secondary
  nist_ai_rmf:
    - subcategory: "MG.2.3"
      context: "Treating symlink-based sandbox escape as an identified AI risk requires active runtime countermeasures; this detection rule is the primary risk treatment implementation."
      strength: primary
    - subcategory: "MP.5.1"
      context: "Identifying the symlink-to-sensitive-target pattern as an AI risk to be catalogued in the organizational risk register."
      strength: secondary
  iso_42001:
    - clause: "8.1"
      context: "ISO/IEC 42001 Clause 8.1 (operational planning and control) is operationalised by this rule's detection of symlink-based sandbox-escape attempts."
      strength: primary
    - clause: "8.3"
      context: "ISO/IEC 42001 Clause 8.3 (AI risk treatment) is implemented via this rule's runtime detection of the symlink primitive."
      strength: secondary

tags:
  category: privilege-escalation
  subcategory: symlink-sandbox-escape
  scan_target: llm_io
  confidence: high

agent_source:
  type: llm_io
  framework:
    - any
  provider:
    - any

detection:
  condition: any
  conditions:
    - field: content
      operator: regex
      value: '(?i)\bln\s+-[a-z]{1,4}\s+["''\x60]?(?:~|/home/[\w.-]+|/Users/[\w.-]+|/root)(?:/|\\)\.(?:ssh|aws|gnupg|docker|kube|npmrc|netrc)\b(?=\s|["''\x60]|$)'
      description: "ln -s command whose target is exactly a well-known credential directory (.ssh, .aws, .gnupg, .docker, .kube) under a home directory -- the trailing lookahead requires the path to end right there (whitespace/quote/end-of-string), not continue into a subpath, so linking a specific harmless file living inside that directory (e.g. a public key) does not fire here"
    - field: content
      operator: regex
      value: '(?i)\bln\s+-[a-z]{1,4}\s+["''\x60]?\S{0,40}?(?:/etc/(?:shadow|passwd|sudoers)\b|\bid_(?:rsa|ed25519|ecdsa|dsa)(?!\.pub\b)\b|\bauthorized_keys\b|\.aws[/\\]credentials\b)'
      description: 'ln -s command whose target is a system account file or SSH PRIVATE key / authorized_keys file -- the negative lookahead (?!\.pub\b) immediately after the key-name alternation excludes id_rsa.pub/id_ed25519.pub/etc. (public keys), since \b alone treats the "." before "pub" as a word boundary and would otherwise match inside the harmless .pub filename'
  false_positives:
    - "A symlink to an ordinary project file or sibling directory within the repository (e.g. ln -s ../shared/lib ./vendor/lib, ln -s node_modules/.bin/eslint ./bin/eslint) -- no sensitive credential path is referenced"
    - "A symlink to a system library for legitimate build/runtime linking (e.g. ln -s /usr/lib/libssl.so ./vendor/libssl.so) -- not in the sensitive credential/account-file list"
    - "Documentation or a security advisory discussing this symlink technique in prose without an actual ln -s command being issued"
    - "A symlink to a PUBLIC key file inside a credential directory (e.g. ln -s ~/.ssh/id_rsa.pub ./shared/id_rsa.pub) -- routine and harmless since a public key is meant to be shared; excluded by both the directory-target end-anchor (condition 1 requires the .ssh path to terminate there, not continue into a subpath) and the id_rsa negative lookahead (condition 2 excludes any .pub-suffixed key name)"

response:
  actions:
    - block_input
    - alert
  message_template: >
    [ATR-2026-02301] A symlink is being created pointing to a sensitive
    credential path outside the workspace -- treat as a sandbox-escape
    staging step, not routine file linking.

confidence: 76
wild_fp_rate: 0

test_cases:
  true_positives:
    - input: "ln -s /home/victim/.ssh ./workspace/output_dir"
      expected: triggered
      description: "CVE-2026-39861 pattern - symlink from workspace-relative dir to the victim's .ssh credential directory"
    - input: "ln -sf ~/.aws ./scratch/creds_link"
      expected: triggered
      description: "Symlink targeting the home-relative AWS credentials directory"
    - input: "ln -s /etc/shadow ./workspace/tmp/output.log"
      expected: triggered
      description: "Symlink targeting /etc/shadow for later read-through-write exfiltration"
    - input: "ln -sfn ~/.ssh/id_rsa ./workspace/notes/key_backup.txt"
      expected: triggered
      description: "Symlink targeting an SSH private key file directly"
  true_negatives:
    - input: "ln -s ../shared/lib ./vendor/lib"
      expected: not_triggered
      description: "Ordinary symlink to a sibling project directory, no credential path referenced"
    - input: "ln -s node_modules/.bin/eslint ./bin/eslint"
      expected: not_triggered
      description: "Ordinary build-tooling symlink into node_modules, no credential path referenced"
    - input: "ln -s /usr/lib/libssl.so ./vendor/libssl.so"
      expected: not_triggered
      description: "Ordinary system-library symlink for build linking, not a sensitive credential/account path"
    - input: "The advisory describes how ln -s can be used to point a workspace file at a location outside the sandbox."
      expected: not_triggered
      description: "Documentation/advisory prose describing the technique, no actual ln -s command issued"
    - input: "ln -s ~/.ssh/id_rsa.pub ./shared/id_rsa.pub"
      expected: not_triggered
      description: "Regression: adversarial review confirmed this firing -- symlinking a PUBLIC key is routine and harmless. Previously matched via TWO independent bugs: condition 1's .ssh-directory pattern lacked an end-of-path anchor so it matched any subfile under .ssh (not just the bare directory), and condition 2's \\bid_rsa\\b matched inside id_rsa.pub because the boundary before '.pub' still counts as a word boundary. Fixed both: condition 1 now requires the .ssh path to terminate right there, and condition 2 excludes .pub-suffixed matches via a negative lookahead."