Botgate

Botgate shows what a Web Bot Auth signature actually protects—and tests whether your server enforces the boundary you intended.

If a signed request for /orders/123 is changed to /orders/999, is the bot still authenticated? See the difference in two self-contained commands:

brew install kraftaa/tap/botgate
botgate demo weak
botgate demo strict

The weak demo signs only the authority:

Signed boundary
  authority  ✓
  method     ○
  path       ○
  query      ○

changed_method  still_valid  authenticated    allowed  PASS
changed_path    still_valid  authenticated    denied   PASS
changed_query   still_valid  authenticated    allowed  PASS

Those results pass because the demo intentionally declares that weak boundary. Botgate does not call a standards-compliant signature a vulnerability merely because it protects less than you expected. It shows the boundary explicitly, then checks it against the policy and access behavior you configured.

The strict demo signs authority, method, path, and query:

changed_method  invalid  unauthenticated  denied  PASS
changed_path    invalid  unauthenticated  denied  PASS
changed_query   invalid  unauthenticated  denied  PASS

Botgate keeps five questions separate:

1. Is the message shaped correctly for a selected protocol profile?
2. Does the signature verify with a supplied or safely discovered key?
3. Which request properties are cryptographically bound?
4. Does that coverage satisfy the application's declared policy?
5. Does an authorized server authenticate the identity, and does the application independently allow access?

Signature validity, authenticated identity, and application access are not the same result. Botgate deliberately does not infer authentication from an HTTP status code. Live tests use a dedicated authentication signal, normally a response header, and can measure application access through a separate oracle.

Status

The current implementation includes:

• current draft-ietf-webbotauth-httpsig-protocol-00 inspection;
• the current Cloudflare compatibility profile;
• current dictionary-form and legacy string-form Signature-Agent handling;
• RFC 9421 signature-base reconstruction for the common request components;
• RFC 8941 canonicalization for selected Structured Field dictionary members;
• Ed25519 signing and verification;
• JWK thumbprint/key selection;
• semantic coverage and configurable policy evaluation;
• Content-Digest SHA-256 validation;
• JSON output for CI;
• an offline mutation matrix;
• controlled live-server testing with independent authentication and access oracles;
• expired, future-created, long/missing-expiry, unknown-key, method, path, query, authority, covered-header, identity-header, and corrupted-signature cases;
• SSRF-resistant Signature-Agent discovery for directory, jwks_uri, and CIMD forms;
• a local well-known key-directory server;
• self-contained strict, weak-boundary, and access-denial demonstrations;
• a local verifier for custom reproducible demonstrations.

Network operations disable redirects and proxies, pin a validated DNS result, cap response sizes and timeouts, and reject private, loopback, link-local, and special-use addresses by default. Local test overrides are explicit.

Install

You do not need Rust or Cargo to use Botgate.

On macOS or Linux, install with Homebrew:

brew install kraftaa/tap/botgate

If you already use Python tooling, install the same native Botgate binary on macOS, Linux, or Windows with pipx:

pipx install botgate

Or with uv:

uv tool install botgate

Inside an existing Python virtual environment, ordinary pip works too:

python -m pip install botgate

These package-manager installs use prebuilt native binaries; they do not compile Botgate or install a Rust toolchain. Standalone archives, a Windows installer, a shell installer, and SHA-256 checksums are also available on the Releases page as fallbacks.

Confirm the installation:

botgate --version
botgate --help

Built-in demonstrations

The built-in scenarios create temporary keys, start an isolated loopback verifier on an available port, run the mutation matrix, and clean up automatically. They need no configuration files and send nothing outside the machine.

Start with the weak-boundary scenario:

botgate demo weak

It signs only the authority. Method, path, and query mutations therefore remain cryptographically valid and authenticated. Application access independently denies the changed path while allowing the changed method and query. These are passing observations under the demo's deliberately weak declared policy—not automatic policy failures.

Compare it with strict binding:

botgate demo strict

Here method, path, and query are signed. Changing any of them invalidates authentication and protected access is denied.

Finally, demonstrate authorization remaining independent of authentication:

botgate demo access

The original request has a valid signature and authenticated identity, but access to the admin resource is denied. Each scenario also supports --format json for machine-readable output.

Manual end-to-end example

Generate a test key and policy:

botgate init

This creates:

.botgate/
├── private.key       # Ed25519 PKCS#8, mode 0600 on Unix
├── public.jwk
├── directory.json
└── botgate.toml

The whole .botgate/ directory is excluded by .gitignore. Botgate never prints the private key. A conformant remote key directory must be served over HTTPS.

Running botgate init --force regenerates public.jwk and directory.json from the existing private key. It never overwrites private.key or an existing botgate.toml.

On Unix, the private key is created atomically with mode 0600. On Windows it is created with create_new but inherits the directory's ACL; restrict access to .botgate/ yourself.

Sign the example request using the current IETF form:

botgate sign examples/request.http \
  --agent https://agent.example \
  --components @authority,@method,@path,@query \
  --output signed-request.http

Inspect declared coverage without loading a key:

botgate inspect signed-request.http \
  --config examples/strict-policy.toml

Verify cryptographically using a local JWKS:

botgate verify signed-request.http \
  --jwks .botgate/directory.json \
  --config examples/strict-policy.toml

Show the offline mutation matrix (a request file means nothing is sent):

botgate test signed-request.http \
  --jwks .botgate/directory.json \
  --config examples/strict-policy.toml

The matrix is generated independently for every signature in the request.

Custom live end-to-end test

For custom policies and requests, start the low-level local verifier in one terminal:

botgate demo serve \
  --jwks .botgate/directory.json \
  --config examples/strict-policy.toml

Then sign and test the URL directly:

botgate test 'http://127.0.0.1:8080/orders/123?view=summary' \
  --agent https://agent.example \
  --config examples/live-test.toml

The explicit local-network and HTTP permissions are in examples/live-test.toml. Against the strict verifier, the original request must authenticate and invalid mutations must not authenticate. The demo returns HTTP 200 for both outcomes: its X-Botgate-Authenticated header is the authentication evidence, while X-Botgate-Access independently reports application access. An absent or unrecognized oracle value is indeterminate, not success.

The original weak-versus-strict demonstration is also included. Start the demo with examples/weak-policy.toml, then test with --components @authority --config examples/weak-live-test.toml. Method, path, and query mutations remain cryptographically valid and continue to authenticate; Botgate reports that as conformant behavior matching the deliberately weak signature policy, not as a vulnerability.

To test a real authorized HTTPS endpoint, private-network and HTTP overrides are unnecessary:

botgate test 'https://staging.example.com/protected' \
  --agent https://keys.agent.example \
  --auth-header X-Agent-Authenticated \
  --authenticated-value true \
  --unauthenticated-value false

Botgate generates only GET and HEAD mutations by default. Replaying another method requires --allow-unsafe-methods.

Key discovery and directory serving

Verify a saved request using the key source covered by Signature-Agent:

botgate verify signed-request.http --discover

Discovery requires HTTPS, never follows redirects, validates directory content, limits response bytes and key count, and keeps key material scoped to one Signature-Agent identity. For local development, serve the generated directory at the draft's well-known path:

botgate directory serve --port 8787

The built-in directory server binds to loopback and uses HTTP, so it is for local tests only. A conformant public directory must use HTTPS and the specified media type.

Cloudflare interoperability

The September 2026 working-group draft requires new senders to emit dictionary-form Signature-Agent, keyed by the signature label:

Signature-Agent: sig1="https://agent.example"
Signature-Input: sig1=("@authority" "signature-agent";key="sig1");...

Cloudflare currently documents the older bare-string form. Generate and analyze that form explicitly:

botgate sign examples/request.http \
  --agent https://agent.example \
  --components @authority \
  --legacy-agent \
  --output cloudflare-request.http

botgate inspect cloudflare-request.http --profile cloudflare-2026-10
botgate inspect cloudflare-request.http --profile ietf-draft-00

The dated Cloudflare profile is a snapshot of Cloudflare's published behavior, not a promise about future behavior. The shorter --profile cloudflare remains an accepted compatibility alias. Reports include the profile snapshot date and source. The second command above reports why the same request is legacy rather than conformant for a new IETF-draft sender.

Policy

Policy requirements are application requirements, not claims that every conforming Web Bot Auth signature must cover the same fields. Without --config, Botgate automatically loads .botgate/botgate.toml from the current directory when it exists; otherwise it uses the same built-in defaults produced by init. Every report names the policy it applied (Policy: in text, policy_source in JSON). In CI, pass --config explicitly so a change to .botgate/ cannot silently alter the gate. Unknown keys and negative time limits are rejected.

[policy]
require_authority = true
require_method = true
require_path = true
query = "required"
body = "ignore"
max_age_seconds = 300
max_lifetime_seconds = 300
allowed_future_skew_seconds = 30
require_nonce = false

[target]
url = "https://staging.example.com/protected"
timeout_seconds = 10
max_response_bytes = 1048576
allow_private = false
allow_http = false

[expect]
header = "X-Agent-Authenticated"
authenticated_value = "true"
unauthenticated_value = "false"

# Access is a separate application decision. Omit this table when the
# endpoint exposes authentication but not a reliable access signal.
[expect.access]
header = "X-Access-Granted"
allowed_value = "true"
denied_value = "false"

# Only listed cases are asserted. Other cases still report observed access.
[expect.access.cases]
original = "allowed"
changed_path = "denied"

[tests]
changed_method = true
changed_path = true
changed_query = true
changed_authority = true
changed_signed_header = true
changed_body = true
corrupted_signature = true
removed_signature_agent = true
changed_signature_agent = true
expired = true
future_created = true
long_expiration = true
missing_expires = true
unknown_key = true

[policy] describes the cryptographic boundary the application requires from a signature. [expect.access.cases] separately describes application access behavior expected from the live endpoint. This permits a result such as “signature valid, identity authenticated, access denied” without treating the denial as an authentication failure.

Coverage is semantic rather than a flat string check. For example, @target-uri satisfies authority, path, and query coverage, while @path does not cover the query. Body integrity requires a covered Content-Digest (the whole field, or its sha-256 member via ;key="sha-256") whose SHA-256 value matches the body bytes. Covering only another member, such as sha-512, does not count. Individual @query-param components are reported as partial evidence and do not satisfy a policy requiring the entire query to be bound. @request-target covers path and query for the raw HTTP/1.1 requests Botgate accepts.

JSON and exit behavior

Use --format json with inspect, verify, or test.

botgate inspect request.http --format json

Exit status is 0 when no error-level findings exist, 1 when conformance, compatibility, crypto, or policy findings fail, and 4 for input or configuration errors detected after argument parsing. Clap uses its conventional status 2 for command-line usage errors. JSON retains independent finding categories and cryptographic/identity states so CI does not need to infer meaning from prose.

Live JSON keeps the stable expected_crypto compatibility field alongside expected_authentication, observed_authentication, expected_access, and observed_access. Some values in that field—such as freshness_invalid and policy_invalid—describe more than signature mathematics, so the text report labels the column Expected validity. A future schema can split signature math, protocol validity, and policy validity without conflating them. A request can remain cryptographically valid and authenticated while application access is denied; Botgate does not collapse those facts into a generic “accepted” result.

Trust model

Verification with --jwks establishes that the supplied key validates the signature and reports identity as key_only. --discover reports a verified identity only when key material was fetched from the request's covered Signature-Agent locator and the signature validates with the selected key.

The signing command derives the thumbprint from the private key and refuses to proceed if --jwk names a different key pair.

Similarly, signing Content-Digest binds the digest header. Botgate separately recomputes the digest before reporting body integrity.

Supported signature components

Botgate reconstructs:

• @method
• @authority
• @scheme
• @path
• @query
• @target-uri
• @request-target
• @query-param;name=...
• ordinary HTTP fields
• dictionary field members selected with ;key=...

Unsupported derived components cause verification to fail explicitly rather than being guessed. Input files use a strict raw HTTP/1.0 or HTTP/1.1 request format; malformed control data, folded headers, and ambiguous authorities are rejected before analysis.

Non-goals

Botgate does not impersonate commercial agents, bypass bot protection, solve CAPTCHAs, crawl sites, score “AI readiness,” or replace authorization and bot-management systems.

Development

Rust and Cargo are required only when building Botgate from source or contributing to the project:

cargo build --release
cargo fmt --check
cargo clippy --all-targets -- -D warnings
cargo test
cargo deny check

The minimum supported Rust version is 1.86. CI runs the test suite on Linux, macOS, and Windows.

cargo test includes property tests for the HTTP and Structured Fields parsers. Coverage-guided fuzz targets live in fuzz/ and need a nightly toolchain and cargo-fuzz:

cargo install cargo-fuzz
cargo +nightly fuzz run parse_request
cargo +nightly fuzz run parse_signatures
cargo +nightly fuzz run signature_base

See SECURITY.md for the trust model and how to report vulnerabilities. Maintainer release setup is documented in RELEASING.md.

The protocol is still an Internet-Draft. Profile-specific behavior is kept separate in the report engine, and botgate protocol identifies the implemented draft.

References

• Web Bot Auth working-group draft 00
• RFC 9421: HTTP Message Signatures
• Cloudflare Web Bot Auth documentation