injection-checking

Warn

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: MEDIUMOBFUSCATIONDATA_EXFILTRATIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [OBFUSCATION]: Detection of homoglyph characters and encoding bypass techniques.
  • Evidence in EXTRA_INJECTION_TYPES.md: Uses confusable Unicode characters in domain names (e.g., аpple.com using Cyrillic 'а') and character variants (e.g., Admin, ⓐdmin, ᵃdmin) to demonstrate filter bypasses.
  • Includes examples of double URL encoding (%252e%252e%252f) and Punycode transformations used to hide malicious payloads from security scanners.
  • [DATA_EXFILTRATION]: Detailed instructions for accessing sensitive system files.
  • Payloads are provided for reading /etc/passwd and other configuration files using Server-Side Includes (SSI), LaTeX macros, and PHP variable modification vulnerabilities.
  • [COMMAND_EXECUTION]: Inclusion of various system command execution payloads.
  • The documentation demonstrates the use of commands such as id, whoami, ls, and shell piping in different execution environments including shell-escape enabled LaTeX and SSI-enabled web servers.
  • [REMOTE_CODE_EXECUTION]: Documentation of primitives for achieving remote code execution.
  • Provides methodologies for achieving RCE through LaTeX \write18 and SSI #exec directives, including chaining commands to write persistent web shells.
  • [INDIRECT_PROMPT_INJECTION]: Identification of untrusted data ingestion attack surfaces.
  • Ingestion points: SKILL.md identifies input flows into HTML, JS, SQL, templates, URL fetchers, and XML parsers.
  • Boundary markers: Absent in instructions.
  • Capability inventory: Includes file-write, command execution, and network request capabilities across the linked injection types.
  • Sanitization: Absent; the skill focuses on bypassing existing sanitization.
  • [DYNAMIC_EXECUTION]: Documentation of techniques for runtime environment manipulation.
  • Describes techniques for overriding application variables using extract() and variable variables ($$variable), which allow for dynamic manipulation of application state and code flow.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 9, 2026, 12:00 PM