awesome-claude-md 65/100

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/ai-ml/comfyui-workflows/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/ai-ml/comfyui-workflows/CLAUDE.md· markdown

113- Never mix SD 1.5 and SDXL models in the same workflow without explicit conversion nodes.

114- Never ignore VRAM warnings; generation failures corrupt partially-written outputs.

115- Never share generated images with embedded metadata containing full prompts in sensitive c

… (8 chars elided on L115)

116

117## Testing

Occurrences

1 occurrence · at L115

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/ai-ml/langfuse-observability/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/ai-ml/langfuse-observability/CLAUDE.md· markdown

162- Never use ClickHouse without backup strategy; columnar data loss requires complete re-inge

… (18 chars elided on L162)

163- Never ignore SDK flush errors; unflushed traces in memory are lost on application crash or

… (9 chars elided on L163)

164- Never share project API keys across environments; separate keys for dev/staging/prod with

… (21 chars elided on L164)

165

166## Testing

Occurrences

1 occurrence · at L164

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/ai-ml/n8n-workflows/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/ai-ml/n8n-workflows/CLAUDE.md· markdown

117- Never disable webhook signature verification for payment or security-critical integrations

… (26 chars elided on L117)

118- Never ignore execution errors; configure Error Trigger workflows for alerting and dead-let

… (19 chars elided on L118)

119- Never share Owner role credentials broadly; use Member roles with project-level permission

… (23 chars elided on L119)

120- Never hardcode API keys or passwords in workflow JSON; always use Credentials store with e

… (10 chars elided on L120)

121- Never skip database backup before major upgrades; workflow definitions and execution histo

… (38 chars elided on L121)

Occurrences

1 occurrence · at L119

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/analytics-product/metabase-analytics/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/analytics-product/metabase-analytics/CLAUDE.md· markdown

191- Never run ad-hoc queries against production OLTP databases without read replica or query t

… (14 chars elided on L191)

192- Never ignore query performance warnings; single slow query can exhaust connection pool and

… (27 chars elided on L192)

193- Never share admin credentials; use service accounts with minimal permissions for API integ

… (23 chars elided on L193)

194

195## Testing

Occurrences

1 occurrence · at L193

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/database-messaging/memcached/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/database-messaging/memcached/CLAUDE.md· markdown

81- Never use Memcached as a primary data store; it is volatile cache only.

82- Never store sensitive data (PII, credentials, payment info) without encryption at applicat

… (10 chars elided on L82)

83- Never share a Memcached cluster between production and non-production environments.

84- Never use the `flush_all` command in production without circuit breaker on application sid

… (2 chars elided on L84)

85- Never store objects larger than 1MB (default slab max); split large objects or use alterna

… (13 chars elided on L85)

Occurrences

1 occurrence · at L83

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/devops-infra/dapr-microservices/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/devops-infra/dapr-microservices/CLAUDE.md· markdown

91- Never bypass Dapr sidecar and call services directly by IP or Kubernetes service DNS.

92- Never store application secrets in component YAML files; use secret store references.

93- Never share state store components between applications without key namespacing.

94- Never assume pub/sub messages arrive exactly once; design all handlers idempotently.

95- Never use actors for stateless or high-throughput parallelizable workloads.

Occurrences

1 occurrence · at L93

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/devops-infra/docker-registry/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/devops-infra/docker-registry/CLAUDE.md· markdown

117- Never allow mutable tags in production repositories; overwritten tags break reproducible d

… (11 chars elided on L117)

118- Never ignore replication failures; stale disaster recovery registry causes extended outage

… (23 chars elided on L118)

119- Never share robot account keys across multiple CI/CD pipelines; compromise affects all lin

… (13 chars elided on L119)

120- Never disable content trust for production namespaces; unsigned images bypass supply chain

… (14 chars elided on L120)

121- Never skip registry backup; database corruption without backup loses all image metadata an

… (14 chars elided on L121)

Occurrences

1 occurrence · at L119

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/devops-infra/gitlab-ci/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/devops-infra/gitlab-ci/CLAUDE.md· markdown

143- Never use shell executor for untrusted code; Docker executor provides process and filesyst

… (13 chars elided on L143)

144- Never ignore pipeline failures in `allow_failure: true` without documented risk acceptance

… (1 chars elided on L144)

145- Never share runner tokens across projects; compromise of one project compromises all using

… (11 chars elided on L145)

146

147## Testing

Occurrences

1 occurrence · at L145

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/devops-infra/podman-containers/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/devops-infra/podman-containers/CLAUDE.md· markdown

93- Never use `--privileged` flag; add specific capabilities with `--cap-add` instead.

94- Never store registry credentials in plain text; use `podman login` with keyring integratio

… (2 chars elided on L94)

95- Never share host PID namespace (`--pid=host`) unless debugging specific kernel issues.

96- Never mount Docker socket into Podman containers; use API communication instead.

97- Never ignore AppArmor/SELinux labels; verify with `podman inspect --format '{{.HostConfig.

… (16 chars elided on L97)

Occurrences

1 occurrence · at L95

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/devops-infra/syncthing-filesync/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/devops-infra/syncthing-filesync/CLAUDE.md· markdown

165

166- Never expose Syncthing GUI (port 8384) directly to the internet without HTTPS reverse prox

… (28 chars elided on L166)

167- Never share `config.xml` or `key.pem` between devices; each device must have unique certif

… (30 chars elided on L167)

168- Never use `sendreceive` folder type for backup-only targets; use `receiveonly` to prevent

… (32 chars elided on L168)

169- Never ignore `.stignore` file maintenance; outdated patterns sync sensitive credentials, b

… (32 chars elided on L169)

Occurrences

1 occurrence · at L167

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/devops-infra/temporal-workflows/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/devops-infra/temporal-workflows/CLAUDE.md· markdown

89- Never ignore `workflow.continueAsNew()` for long-running workflows exceeding history limit

… (2 chars elided on L89)

90- Never use `workflow.sleep()` for intervals under 1 second; use local activities instead.

91- Never share workflow state in global variables; pass via parameters and return values only

… (1 chars elided on L91)

92- Never deploy worker code changes without versioning for in-flight workflows.

93- Never store secrets in workflow history; use encrypted Data Converter with external KMS.

Occurrences

1 occurrence · at L91

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/devops-infra/woodpecker-ci/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/devops-infra/woodpecker-ci/CLAUDE.md· markdown

86- Never hardcode branch names; use `CI_REPO_DEFAULT_BRANCH` and glob patterns for portabilit

… (2 chars elided on L86)

87- Never use privileged mode for build containers; breaks container isolation and exposes hos

… (16 chars elided on L87)

88- Never share Woodpecker secrets across unrelated repositories; compromise cascades to all l

… (15 chars elided on L88)

89- Never run untrusted PR pipelines with Docker socket access; use restricted agent pools for

… (12 chars elided on L89)

90

Occurrences

1 occurrence · at L88

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/frontend/fresh-preact/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/frontend/fresh-preact/CLAUDE.md· markdown

804. Never create islands for components that do not need client-side interactivity. Prefer st

… (32 chars elided on L80)

815. Never use `npm install` or create a `package.json`. Dependencies are managed through `den

… (19 chars elided on L81)

826. Never share state between islands via global variables. Each island hydrates independentl

… (39 chars elided on L82)

83

84## Testing

Occurrences

1 occurrence · at L82

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561897cde7225eed2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/security/authentik-sso/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/security/authentik-sso/CLAUDE.md· markdown

104- Never allow password-based admin access without MFA enforcement; compromise is total compr

… (6 chars elided on L104)

105- Never ignore failed authentication alerts; investigate patterns indicating credential stuf

… (5 chars elided on L105)

106- Never share OAuth client secrets across multiple applications; each app gets unique creden

… (6 chars elided on L106)

107

108## Testing

Occurrences

1 occurrence · at L106

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/security/pomerium-zero-trust/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/security/pomerium-zero-trust/CLAUDE.md· markdown

108- Never use self-signed certificates without pinning; browsers and API clients reject them.

109- Never configure `allow_public_unauthenticated_access` without explicit risk documentation

… (13 chars elided on L109)

110- Never share the same client credentials across multiple IdP configurations.

111- Never ignore certificate expiration alerts; expired proxy cert causes complete access deni

… (3 chars elided on L111)

112- Never store Pomerium signing keys in version control or unencrypted persistent volumes.

Occurrences

1 occurrence · at L110

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/specialized/jellyfin-media/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/specialized/jellyfin-media/CLAUDE.md· markdown

145- Never ignore database corruption warnings; run `PRAGMA integrity_check` on SQLite before m

… (14 chars elided on L145)

146- Never use software transcoding for >3 concurrent 1080p streams without dedicated CPU cores

… (32 chars elided on L146)

147- Never share admin account credentials; create separate user accounts with appropriate libr

… (11 chars elided on L147)

148

149## Testing

Occurrences

1 occurrence · at L147

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/specialized/mattermost-chat/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/specialized/mattermost-chat/CLAUDE.md· markdown

183- Never disable database connection pooling; PostgreSQL default 100 connections; tune `MaxOp

… (24 chars elided on L183)

184- Never ignore Elasticsearch index health yellow/red status; search degradation impacts user

… (21 chars elided on L184)

185- Never share admin or system admin credentials; use role-based access with least privilege

… (21 chars elided on L185)

186- Never skip compliance exports for regulated industries (finance, healthcare); legal hold r

… (33 chars elided on L186)

187- Never run RTCD SFU without TURN server; 30-40% of users behind symmetric NAT cannot connec

… (16 chars elided on L187)

Occurrences

1 occurrence · at L185

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha2561b156a6029264b2c…rubric 365aacaView on GitHub

MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · templates/testing/locust-python/CLAUDE.md

MEDIUM — it fires on intent; whether the agent honors the non-disclosure imperative depends on the host model.

Why it matters

A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.

The exact value spotted

excerpttemplates/testing/locust-python/CLAUDE.md· markdown

69

701. Never use `time.sleep()` in tasks; Locust uses gevent, so use `gevent.sleep()` or rely on

… (12 chars elided on L70)

712. Never share mutable state between users without `gevent.lock.Semaphore`; users run as gre

… (6 chars elided on L71)

723. Never hardcode host URLs in locustfiles; always use `self.host` which comes from `--host`

… (10 chars elided on L72)

734. Never skip response validation; always use `catch_response=True` and check status codes a

… (15 chars elided on L73)

Occurrences

1 occurrence · at L71

How to fix

Remove the non-disclosure imperative, or rescope it so it limits output format, not honesty to the user.

Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.

Framework references

OWASPLLM01 ATLASAML.T0051

Trace & refs

ruleSS-SKILL-INJECT-IMPERATIVE-01sha256bdc6da1843b0ea3d…rubric 365aacaView on GitHub

awesome-claude-md— agent skill

Is awesome-claude-md safe to install?

Audit the pieces. Scan the whole. Decide.