Skip to content

fix(secrets): stop suppressing live credentials; rewrite 11 skills - #10

Merged
lucky-216 merged 1 commit into
mainfrom
chore/skill-library-audit-v2.1
Aug 14, 2026
Merged

lucky-216 merged 1 commit into
mainfrom
chore/skill-library-audit-v2.1

Conversation

@lucky-216

Copy link
Copy Markdown
Collaborator

Three real scanner bugs

The secret-detection checklist declared window: 50 on its placeholder
exclusion and the Exclusion struct had no such field, so the words were
matched against the credential itself. A live Stripe test-mode key is
sk_test_… and contains the substring test, so every one of them was
reported as a known false positive.

Window is now implemented, searching a span either side of the match and
cutting the match out — a window spanning it reproduces the bug it exists to
fix, which the first version of this patch did until the new test caught it.
The word list is split accordingly: placeholder values (your-,
REPLACE_ME) still match the value; documentation markers (example,
sample) match only the surrounding context.

The JWT exclusion carried the base64 of {"alg":"HS256","typ":"JWT"} with
match_type: prefix — the header of essentially every HS256 token, so real
bearer tokens were suppressed. Removed.

Stripe pk_live_ drops from high to medium: Stripe documents it as
publishable and intends it to sit in browser JavaScript, so at high it
fired on every correctly-written checkout page.

Upgrade note: scan_secrets and audit now report findings they
previously suppressed. A repository containing an sk_test_ key or an HS256
JWT may see a gate that was green turn red.

Context-only pointers reached no agent

## Context is stripped from every packaged target, and six skills routed
cross-skill delegations through it — threat-modeling did it for its entire
"who owns each control" paragraph. dev validate now rejects a pointer that
appears only there, in both the backtick and [[wiki]] forms; the second
caught dynamic-verification's handoff to security-regression-tests.

Eleven skills rewritten to 2.0.0

protocol-, ssrf-, websocket-, ml-, threat-modeling, secret-detection,
secure-code-review, template-injection, security-regression-tests,
serverless- and saas-security.

Vendor and version specifics move into references/, where the convention
allows them to date. Claims that could not be verified offline were removed
rather than replaced — saas-security's vendor table went entirely, because
its values disagreed with the rules/*.json this same repo ships.
ml-security's duplicate deserializer rules went too: no code read them, and
the torch.load regex matched every call including the safe one.

Also corrects the claim that secure-vibe audit enforces a committed test.
It is a scanner; it neither runs a suite nor notices a missing one.

The placeholder exclusion in secret_detection.yaml declared `window: 50`
and the Exclusion struct had no such field, so the words were matched
against the credential itself. A live Stripe test-mode key is `sk_test_…`
and contains the substring "test", so every one of them was reported as a
known false positive. Implements Window, searching a span either side of
the match and deliberately cutting the match out — a window spanning it
reproduces the bug it exists to fix. Splits the word list accordingly:
placeholder values ("your-", "REPLACE_ME") still match the value, while
documentation markers ("example", "sample") match only the context.

Removes the JWT exclusion. Its prefix decoded to {"alg":"HS256","typ":"JWT"},
the header of essentially every HS256 token, so `match_type: prefix`
suppressed real bearer tokens. Downgrades Stripe pk_live_ from high:
Stripe documents it as publishable and intends it to sit in browser
JavaScript, so at high it fired on every correct checkout page.

Adds a validator for cross-skill pointers that appear only in `## Context`.
That section reaches no packaged target, so six skills routed delegations
through text the build strips — threat-modeling did it for its entire
"who owns each control" paragraph. Matches both the backtick and the
[[wiki]] pointer forms; the second caught dynamic-verification's handoff
to security-regression-tests.

Rewrites protocol-, ssrf-, websocket-, ml-, threat-modeling,
secret-detection, secure-code-review, template-injection,
security-regression-tests, serverless- and saas-security to 2.0.0. Vendor
and version specifics move into references/ where they are allowed to
date; claims that could not be verified offline were removed rather than
replaced. Drops saas-security's vendor table, whose values disagreed with
the rules/*.json the same repo ships, and ml-security's duplicate
deserializer rules, which no code read and whose torch.load regex matched
every call including the safe one.

Corrects the claim that `secure-vibe audit` enforces a committed test: it
is a scanner and neither runs a suite nor notices a missing one.
@lucky-216
lucky-216 merged commit 634ee02 into main Aug 14, 2026
3 checks passed
@lucky-216
lucky-216 deleted the chore/skill-library-audit-v2.1 branch August 14, 2026 02:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant