diff --git a/docs/integrations/mcp.mdx b/docs/integrations/mcp.mdx index 6b9945c9..35f069a4 100644 --- a/docs/integrations/mcp.mdx +++ b/docs/integrations/mcp.mdx @@ -22,10 +22,17 @@ Create the directory if it does not exist, then write the file: mkdir -p ~/.strix ``` -Paste the servers you want into `~/.strix/mcp-servers.json`. The example below shows one of each transport: a local filesystem server over `stdio` and a remote GitHub server over `http` with a bearer token: +Paste the servers you want into `~/.strix/mcp-servers.json`. The example below shows local `stdio` servers (CyberChef for payload deobfuscation and a local filesystem server) alongside a remote GitHub server over `http` with a bearer token: ```json [ + { + "name": "cyberchef", + "transport": "stdio", + "command": "npx", + "args": ["-y", "@noorfatima123456/cyber-chef-mcp"], + "notes": "CyberChef MCP server for multi-layer payload deobfuscation, crypto decoding, and entropy analysis." + }, { "name": "local_fs", "transport": "stdio", diff --git a/strix/skills/tooling/cyberchef.md b/strix/skills/tooling/cyberchef.md new file mode 100644 index 00000000..d618a161 --- /dev/null +++ b/strix/skills/tooling/cyberchef.md @@ -0,0 +1,195 @@ +--- +name: cyberchef +description: Multi-layer payload deobfuscation, cryptographic decoding, heuristic recipe detection (magic), and entropy analysis via CyberChef MCP. +--- + +# CyberChef MCP Tooling Playbook + +Official resources: +- https://github.com/gchq/CyberChef +- https://github.com/noor202401938-netizen/cyber-chef-mcp +- https://gchq.github.io/CyberChef/ +- https://modelcontextprotocol.io + +CyberChef MCP provides 33 core data transformation, cryptographic, compression (Gunzip, Gzip, Zlib, Raw Deflate), and forensic operations with zero external dependencies. Connected via the Model Context Protocol (MCP) server `cyberchef`, it enables Strix agents to autonomously analyze, deobfuscate, unpack, and verify encoded exploit payloads, authorization tokens, and obfuscated attack vectors with deterministic sub-millisecond execution. + +## MCP Discovery & Dispatch Workflow + +In Strix, agents interact with external MCP servers through the standard generic-dispatch tools: +1. **Discover Connection**: Call `list_mcps()` to verify that the `cyberchef` connection is active. +2. **Search Tools**: Call `search_mcp_tools(connection="cyberchef", query="magic")` to locate matching tool names. +3. **Inspect Schema**: Call `get_mcp_tool_schema(connection="cyberchef", tool="cyberchef_magic")` to review argument parameters. +4. **Dispatch Call**: Call `call_mcp(connection="cyberchef", tool="", arguments={...})` to execute the operation. + +## High-Signal CyberChef Tools + +When connected to `cyberchef`, the following tools are available on the connection: + +- `cyberchef_magic`: Run heuristic detection across known encodings, ciphers, and hash formats. Returns recommended deobfuscation recipes and confidence scores. +- `cyberchef_bake`: Execute sequential operation chains (e.g. `[{"op": "From Base64"}, {"op": "URL Decode"}]`). +- `cyberchef_from_base64`: Decode standard or URL-safe Base64 strings. +- `cyberchef_to_base64`: Encode plaintext into Base64 / URL-safe Base64. +- `cyberchef_from_hex`: Convert hexadecimal sequences to text (supports `None`, `Space`, `0x`, `Comma` delimiters). +- `cyberchef_url_decode`: Decode single or multi-round percent-encoded parameters. +- `cyberchef_rot13`: Rotate characters by offset (default 13, Caesar cipher support). +- `cyberchef_xor`: Decrypt or apply bitwise XOR with secret key. +- `cyberchef_jwt_decode`: Parse and inspect header, claims, alg, and signatures of JSON Web Tokens. +- `cyberchef_entropy`: Calculate Shannon entropy to distinguish plaintext, compressed data, and encrypted or packed payloads. +- `cyberchef_defang_url`: Defang malicious or suspicious indicators (`hxxps://target[.]com`) for safe reporting. +- `cyberchef_extract_entities`: Extract URLs, IP addresses, and email addresses from raw logs or memory strings. + +## Agent-Safe Baseline for Automation + +### 1. Heuristic First (`cyberchef_magic`) +Always run `cyberchef_magic` via `call_mcp` on unknown high-entropy or encoded strings before guessing transformations: +```json +{ + "tool": "call_mcp", + "arguments": { + "connection": "cyberchef", + "tool": "cyberchef_magic", + "arguments": { + "input": "ZXlKaGJHY2lPaUpTVXpVbkxh..." + } + } +} +``` + +### 2. Sequential Multi-Layer Deobfuscation (`cyberchef_bake`) +For payloads with layered obfuscation (e.g. Hex inside Base64 inside URL-encoded query params): +```json +{ + "tool": "call_mcp", + "arguments": { + "connection": "cyberchef", + "tool": "cyberchef_bake", + "arguments": { + "input": "%34%38%36%35%36%63%36%63%36%66", + "recipe": [ + { "op": "URL Decode" }, + { "op": "From Hex", "args": ["None"] } + ] + } + } +} +``` + +### 3. Entropy Assessment & Raw-Byte Workflow +When evaluating whether a payload or parameter is encrypted, packed shellcode, or benign text, evaluate Shannon entropy through `call_mcp`. + +Depending on whether you are assessing an encoded representation directly or require true 8-bit raw-byte entropy, use one of the two workflows below: + +#### Workflow A: Direct Representation-Calibrated Entropy (`cyberchef_entropy`) +Pass the encoded string (Hex or Base64) directly to `cyberchef_entropy` without prior decoding: +```json +{ + "tool": "call_mcp", + "arguments": { + "connection": "cyberchef", + "tool": "cyberchef_entropy", + "arguments": { + "input": "01a2fe89cb994821a0d8e4..." + } + } +} +``` +`cyberchef_entropy` automatically detects the representation alphabet and returns `shannonEntropy`, `bitsPerChar`, `maxForAlphabet`, and `normalizedRatio` (saturation). + +> [!IMPORTANT] +> **Calibrate entropy thresholds by input encoding representation:** +> Shannon entropy measures bits of information per character. The theoretical maximum is strictly bounded by the alphabet size ($\log_2(N)$): +> - **Hex Strings (16 characters, max 4.0 bits/char)**: +> - *Plain text / formatted Hex*: ~2.5 – 3.2 +> - *High-entropy ciphertext / encrypted payload*: **3.8 – 4.0** (Cannot exceed 4.0!) +> - *Warning*: Do not misclassify hex ciphertext scoring ~3.9 as low-entropy content. +> - **Base64 Strings (64 characters, max 6.0 bits/char)**: +> - *Plain text Base64*: ~3.8 – 4.5 +> - *High-entropy ciphertext / packed data*: **5.7 – 6.0** (Cannot exceed 6.0!) +> - **Normalized Saturation Rule**: If `normalizedRatio >= 0.85` (or `verdict == "encrypted_or_compressed"`), the payload is near-maximal entropy for its alphabet, indicating encryption, CSPRNG keys, or compressed data. + +#### Workflow B: Atomic Raw-Byte Entropy via `cyberchef_bake` (Recommended for Raw Binary Streams) +To evaluate true 8-bit Shannon entropy (max 8.0 bits/byte) on an encoded payload (Base64 or Hex), execute the decode operation and entropy analysis **atomically** in a single `cyberchef_bake` recipe: +```json +{ + "tool": "call_mcp", + "arguments": { + "connection": "cyberchef", + "tool": "cyberchef_bake", + "arguments": { + "input": "rLYFVVXj/6ydut5XjZodQFTcIX3qdGy5lS4CBmdY...", + "recipe": [ + { "op": "From Base64" }, + { "op": "Entropy" } + ] + } + } +} +``` +For Hex-encoded payloads: +```json +{ + "tool": "call_mcp", + "arguments": { + "connection": "cyberchef", + "tool": "cyberchef_bake", + "arguments": { + "input": "4a8f1b9c2d3e4f5a6b7c8d9e0f1a2b3c4d5e6f7a8b...", + "recipe": [ + { "op": "From Hex", "args": ["None"] }, + { "op": "Entropy" } + ] + } + } +} +``` +This in-engine pipeline keeps raw decoded byte buffers in memory without serializing non-UTF-8 bytes across the JSON-RPC boundary. The returned entropy report uses the 8-bit raw alphabet (`maxForAlphabet: 8`): +- **Raw Binary / Decoded Byte Streams (256 values, max 8.0 bits/byte)**: + - *Plain text / uncompressed source code*: < 4.5 bits/byte + - *Compressed archives / packed code / encrypted shellcode*: > 7.2 bits/byte + +> [!CAUTION] +> **Do NOT attempt multi-turn raw byte passing across `call_mcp`:** +> `call_mcp` transports inputs and outputs over JSON-RPC strings. Arbitrary 8-bit binary bytes (such as non-printable ciphertext or compressed streams) cannot be safely represented or transported across JSON-RPC without corruption (mojibake, Unicode replacement characters `\uFFFD`, or truncation). Never call `cyberchef_from_base64` or `cyberchef_from_hex` and then attempt to pass the resulting string to `cyberchef_entropy` in a second `call_mcp` call. Always use `cyberchef_bake` to chain decoding and entropy atomically, or evaluate representation-calibrated entropy directly on the encoded string via `cyberchef_entropy`. + +## Common Security Analysis Patterns + +### Pattern 1: Nested WAF Bypass / Obfuscated Injection Vector +When target web applications accept encoded input in parameters or cookies: +1. Extract candidate parameter from HTTP request or response. +2. Call `call_mcp(connection="cyberchef", tool="cyberchef_magic", arguments={"input": candidate})` to determine layers. +3. Call `call_mcp(connection="cyberchef", tool="cyberchef_bake", arguments={"input": candidate, "recipe": [...]})` with the suggested pipeline to recover the plaintext injection string. +4. Verify whether the underlying query contains unsanitized SQLi (`UNION SELECT`), XSS, or SSRF vectors. + +### Pattern 2: JWT Security Inspection +When encountering `Authorization: Bearer ` or session tokens: +1. Call `call_mcp(connection="cyberchef", tool="cyberchef_jwt_decode", arguments={"token": token})`. +2. Inspect the header: check for `alg: "none"`, `alg: "HS256"` with potential asymmetric public key confusion, or empty signatures. +3. Inspect claims: verify expiry timestamps (`exp`), issuer (`iss`), role/privilege elevations, and user identities. + +### Pattern 3: XOR Obfuscation Recovery +When inspecting hardcoded binary strings, PowerShell scripts, or obfuscated malware droppers: +1. Identify probable key length or common plaintext prefix (e.g., `MZ`, `http`, `function`). +2. Run `call_mcp(connection="cyberchef", tool="cyberchef_xor", arguments={"input": data, "key": candidate_key})` iterating candidate keys to extract underlying C2 endpoints or script payloads. + +### Pattern 4: Encrypted / Compressed Payload & Entropy Classification +When verifying whether an unknown string parameter is ciphertext, packed shellcode, or a high-entropy secret token: +1. **Calibrated check**: Call `call_mcp(connection="cyberchef", tool="cyberchef_entropy", arguments={"input": candidate})`. Inspect `normalizedRatio` and `verdict`. If `normalizedRatio >= 0.85`, it is probable ciphertext or compressed data. +2. **Raw-byte verification**: If strict 8-bit thresholds (> 7.2 bits/byte) are required, run `call_mcp(connection="cyberchef", tool="cyberchef_bake", arguments={"input": candidate, "recipe": [{"op": "From Base64"}, {"op": "Entropy"}]})` (or `From Hex`). + +## Critical Correctness Rules + +- **Use `call_mcp` Dispatch**: Never attempt to call CyberChef tools directly as top-level agent tools. Always dispatch through `call_mcp(connection="cyberchef", tool="...", arguments={...})`. +- **Atomic Raw-Byte Entropy Execution**: Never attempt to pass decoded raw binary bytes between separate `call_mcp` calls. JSON-RPC cannot transport arbitrary non-printable bytes without corruption. When evaluating raw byte entropy for Base64 or Hex payloads, always chain `From Base64`/`From Hex` and `Entropy` atomically in a single `cyberchef_bake` recipe. +- **Do Not Guess Encodings**: If a string contains `=, %, 0x` or unexpected symbols, run `cyberchef_magic` first rather than blindly applying base64 or URL decoding. +- **Preserve Raw Inputs**: Keep the original obfuscated string in agent memory/notes alongside the decoded output for accurate proof-of-concept (PoC) reporting. +- **Fail-Safe Fallback**: If an operation fails during `cyberchef_bake`, isolate the failing recipe step and execute individual tools (`cyberchef_from_base64`, `cyberchef_url_decode`) sequentially via `call_mcp`. +- **Safe Defanging**: Always run `cyberchef_defang_url` via `call_mcp` on confirmed malicious or C2 URLs before writing final markdown reports. + +## Failure Recovery + +- If `cyberchef_from_base64` throws a padding error, retry with `urlSafe: true` or inspect whether characters are URL percent-encoded first. +- If `cyberchef_from_hex` produces unprintable characters, check if the input is big-endian or uses custom delimiters (`0x`, `Space`, `,`). +- If `call_mcp` returns an unknown tool error, call `search_mcp_tools(connection="cyberchef", query="...")` to discover the exact tool names registered by the server. + +If uncertain, query web_search with: +`site:gchq.github.io/CyberChef cyberchef `