Rule catalogue
Every rule, in the open.
These 50 rules are read straight from the engine that runs behind the API — the script on this page is the same file the server imports. Each one says what it matches, what happens when it does, how literal it is, and which provision, if any, it is mapped to.
Output Integrity
The output is finished and carries no machinery: no placeholders, leaked prompts, stack traces or destructive commands.
G1-PLACEHOLDER
Unresolved template placeholder
Baseline
review
Template variables or filler text reached the output instead of real values.
- Action
- Review — held until a person approves it.
- Severity
- medium
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Fail the turn and regenerate: a placeholder means the template was never filled in.
Mapped to
A conduct and safety baseline. No statute is cited for this rule.
Catches
Leaves alone
G1-PROMPT-LEAK
System prompt or instruction leakage
OWASP LLM Top 10 LLM07:2025
block
The output exposes the agent's own instructions or raw chat-template tokens.
- Action
- Block — the output must not be sent.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Never return instruction text. Strip template tokens before rendering.
Mapped to
- OWASP LLM Top 10 LLM07:2025 — System Prompt Leakage · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G1-TOOL-LEAK
Raw tool-call payload in output
OWASP LLM Top 10 LLM05:2025
review
Internal tool or function-call markup leaked into user-visible text.
- Action
- Review — held until a person approves it.
- Severity
- medium
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Render tool results, never the call envelope.
Mapped to
- OWASP LLM Top 10 LLM05:2025 — Improper Output Handling · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G1-INTERNAL-ERROR
Stack trace or internal error exposed
MITRE CWE-209
review
Exception output reveals file paths, queries or library internals.
- Action
- Review — held until a person approves it.
- Severity
- medium
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Catch the error and return a generic message; log the detail server-side.
Mapped to
- MITRE CWE-209 — Error message containing sensitive information · MITRE Common Weakness Enumeration (CWE)
- OWASP LLM Top 10 LLM02:2025 — Sensitive Information Disclosure · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G1-DESTRUCTIVE-CMD
Destructive command or query
OWASP LLM Top 10 LLM06:2025
review
The output contains a command that deletes data, wipes disks or force-rewrites history.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Route destructive operations through an approval step; never emit them as ready-to-run text.
Mapped to
- OWASP LLM Top 10 LLM06:2025 — Excessive Agency · OWASP Top 10 for LLM Applications (2025)
- OWASP LLM Top 10 LLM05:2025 — Improper Output Handling · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G1-ACTIVE-MARKUP
Active content in output
OWASP LLM Top 10 LLM05:2025
block
Script tags, event handlers or data-carrying image links that execute when rendered.
- Action
- Block — the output must not be sent.
- Severity
- high
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Escape model output before rendering; block outbound image URLs with query payloads.
Mapped to
- OWASP LLM Top 10 LLM05:2025 — Improper Output Handling · OWASP Top 10 for LLM Applications (2025)
- MITRE CWE-79 — Cross-site scripting · MITRE Common Weakness Enumeration (CWE)
Catches
Leaves alone
G1-UNSAFE-CODE
Unsafe code pattern
OWASP LLM Top 10 LLM05:2025
advisory
Generated code disables TLS verification, deserialises untrusted data or builds SQL by concatenation.
- Action
- Advisory — reported, does not stop delivery.
- Severity
- medium
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English
- What to do
- Have generated code reviewed or linted before it runs.
Mapped to
- OWASP LLM Top 10 LLM05:2025 — Improper Output Handling · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
Security
No credentials, tokens, internal hosts or hijacked instructions leave the boundary.
G2-AWS-KEY
AWS access key
ISO/IEC 27001 A.5.17
block
An AWS access key ID or secret access key appears in the output.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Rotate the key now. Then stop the agent from reading raw credentials.
Mapped to
- ISO/IEC 27001 A.5.17 — Authentication information · ISO/IEC 27001:2022, Annex A
- ISO/IEC 27001 A.8.12 — Data leakage prevention · ISO/IEC 27001:2022, Annex A
- OWASP LLM Top 10 LLM02:2025 — Sensitive Information Disclosure · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G2-API-TOKEN
Vendor API token
ISO/IEC 27001 A.5.17
block
A live-format token for GitHub, Slack, Stripe, OpenAI, Anthropic, Google, SendGrid, Telegram, npm or AgentGate.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Revoke the token at the vendor and remove secrets from the agent's context.
Mapped to
- ISO/IEC 27001 A.5.17 — Authentication information · ISO/IEC 27001:2022, Annex A
- SOC 2 CC6.1 — Logical access security · AICPA Trust Services Criteria
- OWASP LLM Top 10 LLM02:2025 — Sensitive Information Disclosure · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G2-PRIVATE-KEY
Private key block
ISO/IEC 27001 A.8.24
block
A PEM-encoded private key header appears in the output.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Treat the key as compromised and reissue it.
Mapped to
- ISO/IEC 27001 A.8.24 — Use of cryptography · ISO/IEC 27001:2022, Annex A
- OWASP LLM Top 10 LLM02:2025 — Sensitive Information Disclosure · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G2-JWT
JSON Web Token
ISO/IEC 27001 A.5.17
block
A signed session or access token in JWT format.
- Action
- Block — the output must not be sent.
- Severity
- high
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Invalidate the session and keep tokens out of prompts and tool results.
Mapped to
- ISO/IEC 27001 A.5.17 — Authentication information · ISO/IEC 27001:2022, Annex A
- OWASP LLM Top 10 LLM02:2025 — Sensitive Information Disclosure · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G2-CONNECTION-STRING
Credentials in a connection string
ISO/IEC 27001 A.5.17
block
A database, queue or HTTP URL with an embedded username and password.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Rotate the password; hand the agent a scoped proxy instead of a DSN.
Mapped to
- ISO/IEC 27001 A.5.17 — Authentication information · ISO/IEC 27001:2022, Annex A
- PCI DSS Req. 8.3.2 — Authentication factors unreadable in transit and storage · PCI DSS v4.0.1
Catches
Leaves alone
G2-PASSWORD-ASSIGN
Password or secret value
ISO/IEC 27001 A.5.17
block
A password, secret or API key is stated as a literal value.
- Action
- Block — the output must not be sent.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Send credentials through a reset link or secret manager, never in agent text.
Mapped to
- ISO/IEC 27001 A.5.17 — Authentication information · ISO/IEC 27001:2022, Annex A
- SOC 2 CC6.1 — Logical access security · AICPA Trust Services Criteria
Catches
Leaves alone
G2-INTERNAL-HOST
Internal network address
ISO/IEC 27001 A.8.12
advisory
Private IP ranges, cloud metadata endpoints or internal hostnames.
- Action
- Advisory — reported, does not stop delivery.
- Severity
- medium
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English
- What to do
- Replace internal addresses with public documentation links.
Mapped to
- ISO/IEC 27001 A.8.12 — Data leakage prevention · ISO/IEC 27001:2022, Annex A
- OWASP LLM Top 10 LLM02:2025 — Sensitive Information Disclosure · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G2-INJECTION-ECHO
Prompt-injection artifact
OWASP LLM Top 10 LLM01:2025
block
The output repeats override phrases — a sign the agent is echoing or obeying injected instructions.
- Action
- Block — the output must not be sent.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Drop the turn and inspect the retrieved content that carried the injection.
Mapped to
- OWASP LLM Top 10 LLM01:2025 — Prompt Injection · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
Data Protection
Personal, payment and special-category data stay out of the response.
G3-PII-EMAIL
Personal email address
GDPR Art. 5(1)(c)
redact
An individual's email address (role mailboxes such as support@ are ignored).
- Action
- Redact — the value is masked in
redacted_output. - Severity
- high
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Use the redacted output, or reference the customer by an internal ID.
Mapped to
- GDPR Art. 5(1)(c) — Data minimisation · Regulation (EU) 2016/679
- GDPR Art. 5(1)(f) — Integrity and confidentiality · Regulation (EU) 2016/679
- CCPA / CPRA §1798.100(c) — Proportionate use · Cal. Civ. Code §1798.100 et seq.
Catches
Leaves alone
G3-PII-PHONE
Personal phone number
GDPR Art. 5(1)(c)
redact
A phone number in international, North American, Russian or Uzbek format (toll-free lines are ignored).
- Action
- Redact — the value is masked in
redacted_output. - Severity
- medium
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English
- What to do
- Use the redacted output, or show only the last two digits.
Mapped to
- GDPR Art. 5(1)(c) — Data minimisation · Regulation (EU) 2016/679
- GDPR Art. 5(1)(f) — Integrity and confidentiality · Regulation (EU) 2016/679
Catches
Leaves alone
G3-PII-SSN
US Social Security number
CCPA / CPRA §1798.140(ae)
redact
A nine-digit SSN in a valid area/group/serial range.
- Action
- Redact — the value is masked in
redacted_output. - Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Use the redacted output. An SSN should never appear in agent text.
Mapped to
- CCPA / CPRA §1798.140(ae) — Sensitive personal information · Cal. Civ. Code §1798.100 et seq.
- GDPR Art. 5(1)(f) — Integrity and confidentiality · Regulation (EU) 2016/679
Catches
Leaves alone
G3-PII-CARD
Payment card number
PCI DSS Req. 3.4.1
redact
A 13–19 digit number that passes the Luhn check and starts with a card-network prefix (including UzCard 8600 and Humo 9860).
- Action
- Redact — the value is masked in
redacted_output. - Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Use the redacted output — PCI DSS allows at most the BIN and last four digits on display.
Mapped to
- PCI DSS Req. 3.4.1 — PAN masked when displayed · PCI DSS v4.0.1
- PCI DSS Req. 3.5.1 — PAN unreadable wherever stored · PCI DSS v4.0.1
- GDPR Art. 5(1)(f) — Integrity and confidentiality · Regulation (EU) 2016/679
Catches
Leaves alone
G3-PII-CVV
Card verification code
PCI DSS Req. 3.3.1
block
A CVV/CVC value stated next to its label.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English, Russian
- What to do
- Block. Verification codes must not be retained or repeated after authorisation.
Mapped to
- PCI DSS Req. 3.3.1 — Sensitive authentication data not retained after authorisation · PCI DSS v4.0.1
Catches
Leaves alone
G3-PII-IBAN
IBAN
GDPR Art. 5(1)(f)
redact
An International Bank Account Number with a valid country length and mod-97 checksum.
- Action
- Redact — the value is masked in
redacted_output. - Severity
- high
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Use the redacted output, or show the last four characters only.
Mapped to
- GDPR Art. 5(1)(f) — Integrity and confidentiality · Regulation (EU) 2016/679
- GDPR Art. 5(1)(c) — Data minimisation · Regulation (EU) 2016/679
Catches
Leaves alone
G3-PII-GOV-ID
Government ID number
GDPR Art. 87
redact
Uzbekistan PINFL, UK National Insurance number, Russian SNILS, or a passport number stated next to the word passport.
- Action
- Redact — the value is masked in
redacted_output. - Severity
- critical
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian, Uzbek
- What to do
- Use the redacted output. Government identifiers need a documented purpose before disclosure.
Mapped to
- GDPR Art. 87 — National identification numbers · Regulation (EU) 2016/679
- GDPR Art. 5(1)(f) — Integrity and confidentiality · Regulation (EU) 2016/679
- Uzbekistan Personal Data Law ZRU-547 — Personal data processing · Law ZRU-547, as amended by ZRU-1125
Catches
Leaves alone
G3-PII-DOB
Date of birth
GDPR Art. 5(1)(c)
redact
A birth date stated next to an explicit label.
- Action
- Redact — the value is masked in
redacted_output. - Severity
- medium
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English, Russian, Uzbek
- What to do
- Use the redacted output, or confirm identity without repeating the date.
Mapped to
- GDPR Art. 5(1)(c) — Data minimisation · Regulation (EU) 2016/679
Catches
Leaves alone
G3-HEALTH-DATA
Health information about a person
GDPR Art. 9(1)
review
A diagnosis or condition stated about an identifiable individual.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English
- What to do
- Hold for review: health data needs an explicit legal basis before it is disclosed.
Mapped to
- GDPR Art. 9(1) — Special categories of personal data · Regulation (EU) 2016/679
- HIPAA 45 CFR §164.502 — Uses and disclosures of protected health information · HIPAA Privacy Rule, 45 CFR Part 164
Catches
Leaves alone
G3-SPECIAL-CATEGORY
Special-category data about a person
GDPR Art. 9(1)
review
Religion, ethnicity, sexual orientation, political opinion, union membership or criminal record attributed to an individual.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English
- What to do
- Hold for review and remove the attribute unless processing it is strictly necessary and lawful.
Mapped to
- GDPR Art. 9(1) — Special categories of personal data · Regulation (EU) 2016/679
- GDPR Art. 10 — Criminal convictions and offences · Regulation (EU) 2016/679
Catches
Leaves alone
G3-BULK-PII
Bulk personal data
GDPR Art. 5(1)(c)
block
Five or more personal identifiers in a single output — a record dump, not a reply.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- A count across other rules · literal, low false-positive risk
- Languages
- English
- What to do
- Block. Paginate behind an authenticated export instead of returning records in chat.
Mapped to
- GDPR Art. 5(1)(c) — Data minimisation · Regulation (EU) 2016/679
- GDPR Art. 32(1) — Security of processing · Regulation (EU) 2016/679
Catches
Leaves alone
Action Safety
Money movement, irreversible operations and privilege changes wait for a human.
G4-MONEY-MOVEMENT
Money movement without approval
EU AI Act Art. 14
review
The agent states it is moving money at or above the approval threshold (default: about 1,000 USD).
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Hold for review. Pass context.metadata.approved_by once a human has signed off.
Mapped to
- EU AI Act Art. 14 — Human oversight (high-risk systems, from 2 Dec 2027) · Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
- SOX §404 — Internal control over financial reporting · Sarbanes-Oxley Act of 2002
- OWASP LLM Top 10 LLM06:2025 — Excessive Agency · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G4-IRREVERSIBLE
Irreversible operation
EU AI Act Art. 14
review
The agent reports deleting, closing, terminating or revoking something that cannot be undone.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Hold for review before the action runs; require an explicit confirmation step.
Mapped to
- EU AI Act Art. 14 — Human oversight (high-risk systems, from 2 Dec 2027) · Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
- OWASP LLM Top 10 LLM06:2025 — Excessive Agency · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
G4-PRIVILEGE-CHANGE
Privilege or security-control change
ISO/IEC 27001 A.8.2
review
Admin rights granted, MFA or logging switched off, or a resource made public.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English
- What to do
- Hold for review. Privileged changes need a named approver.
Mapped to
- ISO/IEC 27001 A.8.2 — Privileged access rights · ISO/IEC 27001:2022, Annex A
- ISO/IEC 27001 A.5.15 — Access control · ISO/IEC 27001:2022, Annex A
- SOC 2 CC6.3 — Role-based access and least privilege · AICPA Trust Services Criteria
Catches
Leaves alone
G4-EXTERNAL-SHARE
Data sent to an external destination
GDPR Art. 5(1)(f)
review
The agent reports sending or uploading material to a personal account or public service, or a customer dataset to an outside party.
- Action
- Review — held until a person approves it.
- Severity
- medium
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Hold for review and confirm the destination is an approved processor.
Mapped to
- GDPR Art. 5(1)(f) — Integrity and confidentiality · Regulation (EU) 2016/679
- OWASP LLM Top 10 LLM06:2025 — Excessive Agency · OWASP Top 10 for LLM Applications (2025)
Catches
Leaves alone
Legal
No unqualified legal or medical advice, binding promises, absolute claims or threats.
G5-LEGAL-ADVICE
Legal advice without qualification
ABA Model Rules 5.5
review
A definite legal conclusion or instruction with no pointer to a qualified lawyer.
- Action
- Review — held until a person approves it.
- Severity
- medium
- Detection
- A trigger phrase without its required safeguard · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Add the qualification, or hand off to counsel.
Mapped to
- ABA Model Rules 5.5 — Unauthorized practice of law · ABA Model Rules of Professional Conduct
Catches
Leaves alone
G5-MEDICAL-ADVICE
Medical directive without referral
EU MDR Annex VIII, Rule 11
review
A diagnosis or dosing instruction with no referral to a clinician.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- A trigger phrase without its required safeguard · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Add the referral, or route the conversation to a clinician.
Mapped to
- EU MDR Annex VIII, Rule 11 — Software providing information used for diagnostic or therapeutic decisions · Regulation (EU) 2017/745 (medical devices)
Catches
Leaves alone
G5-ABSOLUTE-CLAIM
Absolute or unverifiable claim
UCPD Art. 6(1)
advisory
"100% safe", "risk-free", "clinically proven" and similar claims that cannot be substantiated.
- Action
- Advisory — reported, does not stop delivery.
- Severity
- medium
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Replace with a specific, sourced statement.
Mapped to
- UCPD Art. 6(1) — Misleading actions · Directive 2005/29/EC (unfair commercial practices)
- FTC Act §5(a) — Unfair or deceptive acts or practices · 15 U.S.C. §45
Catches
Leaves alone
G5-BINDING-COMMITMENT
Binding promise made by the agent
Case law Moffatt v. Air Canada, 2024 BCCRT 149
review
The agent guarantees a refund, compensation, waiver or approval on the company's behalf.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Hold for review. Courts have held companies to what their chatbots promised.
Mapped to
- Case law Moffatt v. Air Canada, 2024 BCCRT 149 — Company liable for its chatbot's misstatement · Reported decisions
- UCPD Art. 6(1) — Misleading actions · Directive 2005/29/EC (unfair commercial practices)
Catches
Leaves alone
G5-COERCIVE-THREAT
Coercive or false threat
FDCPA 15 U.S.C. §1692e(4)–(5)
block
Threats of arrest, seizure, lawsuits or reporting used to pressure a person.
- Action
- Block — the output must not be sent.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Block. Collections and enforcement language must come from approved templates.
Mapped to
- FDCPA 15 U.S.C. §1692e(4)–(5) — False threats of arrest or legal action · Fair Debt Collection Practices Act, 15 U.S.C. §1692
- UCPD Art. 8 — Aggressive commercial practices · Directive 2005/29/EC (unfair commercial practices)
Catches
Leaves alone
Financial Conduct
No guaranteed returns, unsuitable recommendations or help evading controls.
G6-GUARANTEED-RETURN
Guaranteed or risk-free return
MiFID II Art. 24(3)
block
Investment returns presented as certain, or loss presented as impossible.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English, Russian, Uzbek
- What to do
- Block. No financial promotion may promise a return.
Mapped to
- MiFID II Art. 24(3) — Fair, clear and not misleading · Directive 2014/65/EU and Delegated Regulation (EU) 2017/565
- FCA Handbook COBS 4.2.1R — Fair, clear and not misleading communications · FCA Conduct of Business Sourcebook (COBS)
- FINRA Rule 2210(d)(1)(B) — No promissory or exaggerated claims · FINRA Rules
Catches
Leaves alone
G6-INVESTMENT-ADVICE
Personal investment recommendation without risk warning
MiFID II Art. 25(2)
review
A direct buy/sell instruction with no suitability check or risk warning.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- A trigger phrase without its required safeguard · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Add the risk warning and suitability caveat, or hand off to a licensed adviser.
Mapped to
- MiFID II Art. 25(2) — Suitability of personal recommendations · Directive 2014/65/EU and Delegated Regulation (EU) 2017/565
- SEC 17 CFR §240.15l-1 — Regulation Best Interest · U.S. Securities and Exchange Commission rules
Catches
Leaves alone
G6-PAST-PERFORMANCE
Past performance without warning
MiFID II Delegated Reg. (EU) 2017/565, Art. 44(4)
advisory
A historical return is quoted for an investment with no past-performance warning.
- Action
- Advisory — reported, does not stop delivery.
- Severity
- medium
- Detection
- A trigger phrase without its required safeguard · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Append the standard past-performance warning.
Mapped to
- MiFID II Delegated Reg. (EU) 2017/565, Art. 44(4) — Past performance warnings · Directive 2014/65/EU and Delegated Regulation (EU) 2017/565
Catches
Leaves alone
G6-AML-EVASION
Help evading AML or reporting controls
AML 31 U.S.C. §5324
block
Structuring advice, KYC bypass or laundering instructions.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Block and escalate to the MLRO or compliance officer.
Mapped to
- AML 31 U.S.C. §5324 — Structuring transactions to evade reporting · Directive (EU) 2018/1673 · 31 U.S.C. §5324
- AML Directive (EU) 2018/1673, Art. 4 — Aiding, abetting, inciting and attempting · Directive (EU) 2018/1673 · 31 U.S.C. §5324
Catches
Leaves alone
G6-TAX-EVASION
Tax-evasion advice
US tax code 26 U.S.C. §7201
block
Advice to hide income, skip declarations or fabricate documents.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Block. Offer a referral to a licensed tax adviser.
Mapped to
- US tax code 26 U.S.C. §7201 — Attempt to evade or defeat tax · Internal Revenue Code, 26 U.S.C.
Catches
Leaves alone
Explainability
Adverse decisions come with a reason and a route to a human.
G7-DECISION-NO-REASON
Adverse decision without a reason
GDPR Art. 22(3)
review
A denial, rejection or termination is communicated with no stated reason.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- A trigger phrase without its required safeguard · heuristic — expect some false positives
- Languages
- English, Russian, Uzbek
- What to do
- State the principal reason, or hold the message for a human to complete.
Mapped to
- GDPR Art. 22(3) — Safeguards for automated decisions · Regulation (EU) 2016/679
- ECOA / Reg B 12 CFR §1002.9(a)(2) — Statement of specific reasons for adverse action · 15 U.S.C. §1691 · 12 CFR Part 1002
- EU AI Act Art. 86 — Right to explanation (with the high-risk regime, from 2 Dec 2027) · Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
Catches
Leaves alone
G7-NO-RECOURSE
Adverse decision without a route to a human
GDPR Art. 22(3)
review
A denial is communicated with no way to appeal, contest it or reach a person.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- A trigger phrase without its required safeguard · heuristic — expect some false positives
- Languages
- English, Russian, Uzbek
- What to do
- Add the appeal or contact route before sending.
Mapped to
- GDPR Art. 22(3) — Right to obtain human intervention and to contest the decision · Regulation (EU) 2016/679
- Uzbekistan AI rules Informatization Law, Art. 7¹ — Legally significant decisions may not rest solely on AI conclusions · Law ZRU-1115 of 21 Jan 2026 (Informatization Law, Art. 7¹)
Catches
Leaves alone
G7-LEGAL-CITATION
Legal citation to verify
Case law Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023)
advisory
A case or reporter citation appears; generated citations are a known fabrication risk.
- Action
- Advisory — reported, does not stop delivery.
- Severity
- medium
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English
- What to do
- Verify the citation against a primary source before it is relied on.
Mapped to
- Case law Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023) — Sanctions for fabricated AI-generated citations · Reported decisions
- FRCP Rule 11(b) — Representations to the court · Federal Rules of Civil Procedure
Catches
Leaves alone
Ethics & Transparency
No posing as a human, protected-attribute decisions, manipulation or abuse.
G8-HUMAN-IMPERSONATION
AI claiming to be human
EU AI Act Art. 50(1)
block
The agent says it is a person, or denies being a bot.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English, Russian, Uzbek
- What to do
- Block. People must be told when they are interacting with an AI system.
Mapped to
- EU AI Act Art. 50(1) — Disclosure of AI interaction (applies since 2 Aug 2026) · Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
- California BOT Act §17941 — Bots must not mislead about artificial identity · Cal. Bus. & Prof. Code §17940 et seq.
- Kazakhstan AI Law No. 230-VIII — Duty to inform users of AI interaction · Law on Artificial Intelligence, No. 230-VIII of 17 Nov 2025
Catches
Leaves alone
G8-AI-DISCLOSURE
Missing AI disclosure opt-in
EU AI Act Art. 50(1)
review
Opt-in: when context.metadata.require_ai_disclosure is true, the output must say it comes from an AI system.
- Action
- Review — held until a person approves it.
- Severity
- medium
- Detection
- A required statement is missing · literal, low false-positive risk
- Languages
- English, Russian
- What to do
- Add the disclosure to the first message of the conversation.
Mapped to
- EU AI Act Art. 50(1) — Disclosure of AI interaction (applies since 2 Aug 2026) · Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
- Kazakhstan AI Law No. 230-VIII — Duty to inform users of AI interaction · Law on Artificial Intelligence, No. 230-VIII of 17 Nov 2025
Catches
Leaves alone
G8-PROTECTED-GROUND
Decision grounded on a protected characteristic
GDPR Art. 22(4)
block
A decision is justified by age, sex, religion, ethnicity, disability, pregnancy or similar.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Block and escalate — this is the pattern discrimination claims are built on.
Mapped to
- GDPR Art. 22(4) — Automated decisions must not rest on special categories · Regulation (EU) 2016/679
- ECOA / Reg B 15 U.S.C. §1691(a) — Prohibited bases for credit decisions · 15 U.S.C. §1691 · 12 CFR Part 1002
- EU Charter Art. 21 — Non-discrimination · Charter of Fundamental Rights of the European Union
Catches
Leaves alone
G8-AUTOMATED-DECISION
Solely automated decision
GDPR Art. 22(1)
review
The output says a decision about a person was made automatically, with no human involved.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian, Uzbek
- What to do
- Hold for review, or pass context.metadata.human_reviewed once a person has confirmed it.
Mapped to
- GDPR Art. 22(1) — Right not to be subject to solely automated decisions · Regulation (EU) 2016/679
- Uzbekistan AI rules Informatization Law, Art. 7¹ — Legally significant decisions may not rest solely on AI conclusions · Law ZRU-1115 of 21 Jan 2026 (Informatization Law, Art. 7¹)
- EU AI Act Art. 14 — Human oversight (high-risk systems, from 2 Dec 2027) · Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
Catches
Leaves alone
G8-MANIPULATION
Manipulative pressure
EU AI Act Art. 5(1)(a)–(b)
review
False urgency, invented scarcity, guilt or secrecy used to push a decision.
- Action
- Review — held until a person approves it.
- Severity
- high
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Hold for review; rewrite without the pressure tactic.
Mapped to
- EU AI Act Art. 5(1)(a)–(b) — Prohibited manipulative and exploitative techniques (applies since 2 Feb 2025) · Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
- UCPD Annex I, point 7 — Falsely stating an offer is available for a very limited time · Directive 2005/29/EC (unfair commercial practices)
- DSA Art. 25 — Deceptive interface design · Regulation (EU) 2022/2065
Catches
Leaves alone
G8-ABUSIVE-LANGUAGE
Abusive language
Baseline
review
Insults or profanity directed at the person.
- Action
- Review — held until a person approves it.
- Severity
- medium
- Detection
- Pattern match · literal, low false-positive risk
- Languages
- English, Russian
- What to do
- Hold for review; regenerate with a neutral tone.
Mapped to
A conduct and safety baseline. No statute is cited for this rule.
Catches
Leaves alone
G8-SELF-HARM
Self-harm encouragement
Baseline
block
Content that encourages or instructs self-harm.
- Action
- Block — the output must not be sent.
- Severity
- critical
- Detection
- Pattern match · heuristic — expect some false positives
- Languages
- English, Russian
- What to do
- Block and route to your crisis protocol.
Mapped to
A conduct and safety baseline. No statute is cited for this rule.
Catches
Leaves alone
See them fire on your own text.
The live gate on the home page runs all 50 rules in your browser. Nothing you type is sent.