Trust and security

A conforming host SHOULD expose TrustView (digests, seals, issuer class) before any execute path. Skills are portable; trust decisions MUST travel with the package, not with the originating app.

Practice

1. Inspect before run — digests and seals without executing the skill or feeding package bodies to a model.
2. Validate structure and hash integrity.
3. Dry-run before execute.
4. Treat unsigned / open packages as untrusted.
5. Treat the reference public-dev HMAC mint as development-only — never production identity proof.

skill inspect ./file.skill --trust
skill validate ./file.skill
skill verify-trust ./file.skill
skill run ./file.skill                    # dry-run default

TrustView states

Seal situation	TrustView	Production execute
Unsigned / open	untrusted	Refuse unless --allow-untrusted
Public-dev HMAC	development	Refuse (forgeable)
Configured key + self-reported host	self_reported	Refuse unless opted in
Configured key + verified host binding	verified_issuer	Allowed

Trust profiles (protocol): open | minted | anchored | issuer:<id>.

The trust ladder (mint-time provenance)

TrustView states above answer "does the runtime's execute gate trust this signature." A separate, complementary axis is how much publicly-checkable provenance a package carries at mint time:

Rung	How it's sealed	What a verifier gets
Development	Public dev HMAC key (default, zero setup)	Local iteration only. Forgeable by design, labeled development everywhere it appears, never production trust.
Verified issuer	Configured Ed25519 key (skill keygen + --signer-key)	Cryptographic proof of authorship and integrity, once a verifier pins your key in their trust store.
Publicly anchored	Rekor transparency log (--transparency) and/or Fulcio keyless OIDC (--keyless)	A public, independently-checkable record, anyone can confirm the entry on Sigstore's own infrastructure.

Anchoring is orthogonal to TrustView state and always additive, an anchored package can still be development or self_reported trust; the anchor never replaces the seal. Inclusion is not endorsement: logging a package proves auditability, not goodness.

skill publish makes the publicly-anchored rung frictionless: the public Rekor log needs a signing key but no login, so a per-user key is auto-provisioned on first use. That key alone is not verified_issuer, without real agent-runtime evidence the seal binds self_reported, and third parties earn verified_issuer for your packages only once they pin your key. The public URL works either way; it never inflates what the seal claims.

The PermanenceAnchor slot this ladder's third rung uses is an open extension point: the wire format already reserves a ledger anchor kind alongside the shipped ones, a documented, unimplemented roadmap item, never required, always additive if it ships. skillerr does not mint tokens, issue NFTs, or move value; "minting" a .skill creates a cryptographic attestation, not a financial instrument.

Seals and digests

• Digests: sha256:<hex>
• package_digest covers package content (excludes skill.json and signatures/**)
• sealed_manifest_digest binds title, intent, permissions, policy, capabilities, input sensitivity, content digests, and contract summary — not only workflow / knowledge bytes
• Creation attestation records agent host / provider / model, host_claim_binding (self_reported | verified_issuer), and issuer_class (public_dev_hmac | configured_hmac | verified_issuer)

Inspect without side effects:

skill inspect ./file.skill --trust

SKILL_HOST and anti-spoof

• Mint refuses denylisted hosts: human, cli, shell, manual, bash, terminal, …
• Exporting SKILL_HOST=cursor alone cannot produce verified_issuer trust
• Agent runtime markers (SKILL_AGENT_INVOCATION, SKILL_SESSION_ID, IDE markers) strengthen the mint path but remain locally spoofable
• Seals record claim binding and issuer class so TrustView can distinguish self-reported vs verified issuer

Runtime deny-by-default

• Undeclared network / filesystem / secret capabilities are refused
• Missing consent fails closed for side-effecting steps
• execute / resume refuse unsigned, open, public-dev HMAC, and self_reported seals unless --allow-untrusted

skill run ./file.skill --mode execute --allow-untrusted   # explicit unsafe opt-in

Secrets

Never embed API keys, tokens, .env, or private customer data in sections or journey text. Use / env refs. Continuity and release packages default to shareable-redacted journey content.

Residual risk (honest)

A seal proves which key signed which claims. It cannot prove that a named local LLM was honest about authorship. Treat host / provider / model fields as claims under that key’s honesty — especially offline and self-hosted models.

The bundled development HMAC signer (issuer_class=public_dev_hmac) is never production trust.

Threats to keep in mind

Malicious packages, prompt injection via resources, tool escalation, dependency confusion (including similarly named npm packages). Prefer verifying package identity and digests before run.

Reporting

For the skillerr reference implementation, use private vulnerability reporting via skillerr.com. Do not open public issues with exploit details.