Skip to content

Security: krnlbash/wired2ai

Security

SECURITY.md

Security model

Challenge artifacts and HTTP responses are hostile input. Native tools are trusted executables installed by the host administrator; malformed input can still exploit bugs in those tools. The same applies to Python and its parsers.

Implemented boundaries

  • File input must be a regular file. The final path component cannot be a symlink. Parent-directory symlinks are resolved by the OS; this is not a filesystem jail.
  • A complete, size-limited snapshot is hashed and analyzed. Normal metadata changes during the read are rejected; this does not defeat a malicious writer that can manipulate the underlying filesystem or guarantee forensic immutability.
  • Native subprocesses receive argv arrays, closed stdin, a clean environment, a private working directory, time limits and combined output caps. A fixed system PATH is used. No challenge executable is passed as the executable.
  • The runner kills its process group on completion or limit failure. An exploited tool could deliberately escape that group. Memory, CPU, filesystem and network access are not restricted by a kernel sandbox or cgroup.
  • No target execution, ldd, archive extraction, decompilation plugins, shell command strings, Python eval, authentication flow or automatic LLM upload.
  • Web collection allows HTTP(S) only, verifies TLS using curl's normal trust store, disables curl configuration/proxy inheritance and enforces origin and directory checks before each request. It does not maintain a cookie session.
  • Private network targets are allowed for local CTFs. DNS is not pinned between requests. Do not expose this CLI as an untrusted multi-user web service.
  • The Markdown export uses dynamic fences and visible control-character escapes. This reduces formatting ambiguity; it does not neutralize model prompt injection.

Data handling

Report files are private to the current account by default. Original artifacts are not included in reports. Temporary copies are deleted on normal exit; forceful termination or a crash can leave temporary files. Deletion is not secure erasure. Run in a disposable encrypted VM if the artifacts require that level of handling.

PEM private-key markers suppress generic byte/string previews. OpenSSL exports only public components of unencrypted keys. Encrypted PEM containers are skipped. Embedded or encoded secrets can evade these markers. Common assignments, token forms, authorization values and private PEM blocks are masked best effort. Environment probe values are omitted wholesale. Source paths, binary strings, EXIF data and other metadata can still contain sensitive information.

Recognized PCAPNG inputs have a separate metadata-only export policy. Generic header/string previews, native format/ExifTool extraction and crypto payload previews are suppressed even with a category override. The built-in parser exports numeric structure and offsets, not packet bodies, text options, name resolution records or decryption-secret values. It counts decryption-secrets blocks and records their declared lengths. This policy also applies when the PCAPNG parser fails or encounters an unsupported version; it is not contingent on recognizing a particular key format. File names, hashes, timestamps and other numeric metadata are still shared. A file with a damaged/missing PCAPNG signature is not covered by this format-specific policy.

PCAPNG traversal and retained details have separate hard caps. Unknown block bodies are skipped after framing checks. Supported structures with inconsistent bounds fail without attempting recovery. No captured packet is replayed, resolved or dissected. The parser runs in-process; these bounds are not an OS sandbox.

PE parsing reads snapshot bytes with bounded RVA/file-offset mapping and hard limits on sections, import descriptors/thunks, exports, names, certificate entries and entropy samples. It does not load DLLs, follow forwarded exports or execute entry points. Invalid or ambiguous mappings fail the affected check. Unlike PCAPNG, generic PE strings and declared import/export names are exported, subject to the existing limits and best-effort redaction. They remain untrusted content. Certificate entries are framed, not cryptographically verified; a framed table does not imply a signed/trusted image. Static mitigation flags do not prove runtime enforcement. Windows execution, loader compatibility and malicious-sample coverage are not established by the synthetic tests.

Evidence records are sanitized and bounded; they are not a forensic evidence archive. Keep original artifacts separately. Reading a source can affect atime.

Reporting a problem

Report reproducible failures with a harmless minimized fixture and the output of wired2ai doctor. Do not post real credentials, private challenge data or working secrets. Once the GitHub repository exists, enable private vulnerability reporting before inviting sensitive reports. No private reporting endpoint has been configured in this local release.

There aren't any published security advisories