Audit & Compliance

Triggerfish मधील प्रत्येक policy decision full context सह logged आहे. कोणतेही exceptions नाहीत, logging disable करणारा "debug mode" नाही, आणि LLM audit records suppress करण्याचा कोणताही मार्ग नाही. हे system ने घेतलेल्या प्रत्येक security decision ची complete, tamper-evident record provide करते.

काय Recorded होते

Audit logging एक fixed rule आहे -- ते नेहमी active असते आणि disable केले जाऊ शकत नाही. प्रत्येक enforcement hook execution एक audit record produce करते:

Field	वर्णन
`timestamp`	Decision कधी घेतला गेला (ISO 8601, UTC)
`hook_type`	कोणता enforcement hook run झाला (`PRE_CONTEXT_INJECTION`, `PRE_TOOL_CALL`, `POST_TOOL_RESPONSE`, `PRE_OUTPUT`, `SECRET_ACCESS`, `SESSION_RESET`, `AGENT_INVOCATION`, `MCP_TOOL_CALL`)
`session_id`	Action झालेले session
`decision`	`ALLOW`, `BLOCK`, किंवा `REDACT`
`reason`	Decision चे Human-readable explanation
`input`	Hook trigger केलेला data किंवा action
`rules_evaluated`	Decision ला reach करण्यासाठी कोणते policy rules checked केले गेले
`taint_before`	Action पूर्वी Session taint level
`taint_after`	Action नंतर Session taint level (changed असल्यास)
`metadata`	Hook type साठी specific additional context

Audit Record Examples

Allowed Output

json

{
  "timestamp": "2025-01-29T10:23:47Z",
  "hook_type": "PRE_OUTPUT",
  "session_id": "sess_456",
  "decision": "ALLOW",
  "reason": "Classification check passed",
  "input": {
    "target_channel": "telegram",
    "recipient": "owner"
  },
  "rules_evaluated": [
    "no_write_down",
    "channel_classification"
  ],
  "taint_before": "INTERNAL",
  "taint_after": "INTERNAL"
}

Blocked Write-Down

json

{
  "timestamp": "2025-01-29T10:24:12Z",
  "hook_type": "PRE_OUTPUT",
  "session_id": "sess_456",
  "decision": "BLOCK",
  "reason": "Session taint (CONFIDENTIAL) exceeds effective classification (PUBLIC)",
  "input": {
    "target_channel": "whatsapp",
    "recipient": "external_user_789",
    "effective_classification": "PUBLIC"
  },
  "rules_evaluated": [
    "no_write_down",
    "channel_classification",
    "recipient_classification"
  ],
  "taint_before": "CONFIDENTIAL",
  "taint_after": "CONFIDENTIAL"
}

Taint Escalation सह Tool Call

json

{
  "timestamp": "2025-01-29T10:23:45Z",
  "hook_type": "POST_TOOL_RESPONSE",
  "session_id": "sess_456",
  "decision": "ALLOW",
  "reason": "Tool response classified and taint updated",
  "input": {
    "tool_name": "salesforce.query_opportunities",
    "response_classification": "CONFIDENTIAL"
  },
  "rules_evaluated": [
    "tool_response_classification",
    "taint_escalation"
  ],
  "taint_before": "PUBLIC",
  "taint_after": "CONFIDENTIAL",
  "metadata": {
    "lineage_id": "lin_789xyz",
    "records_returned": 3
  }
}

Agent Delegation Blocked

json

{
  "timestamp": "2025-01-29T10:25:00Z",
  "hook_type": "AGENT_INVOCATION",
  "session_id": "sess_456",
  "decision": "BLOCK",
  "reason": "Agent ceiling (INTERNAL) below session taint (CONFIDENTIAL)",
  "input": {
    "caller_agent_id": "agent_abc",
    "callee_agent_id": "agent_def",
    "callee_ceiling": "INTERNAL",
    "task": "Generate public summary"
  },
  "rules_evaluated": [
    "delegation_ceiling_check",
    "delegation_allowlist",
    "delegation_depth"
  ],
  "taint_before": "CONFIDENTIAL",
  "taint_after": "CONFIDENTIAL"
}

Audit Trace Capabilities

Audit records चार प्रकारे queried केले जाऊ शकतात, प्रत्येक different compliance आणि forensic need serve करतो.

Forward Trace

प्रश्न: "Salesforce record opp_00123ABC मधील data चे काय झाले?"

एक forward trace data element त्याच्या origin point पासून प्रत्येक transformation, session, आणि output मधून follow करतो. हे answers करतो: हा data कुठे गेला, कोणी पाहिला, आणि तो कधी organization च्या बाहेर पाठवला गेला का?

Origin: salesforce.query_opportunities
  --> lineage_id: lin_789xyz
  --> classification: CONFIDENTIAL
  --> session: sess_456

Transformations:
  --> Extracted fields: name, amount, stage
  --> LLM summarized 3 records into pipeline overview

Outputs:
  --> Sent to owner via Telegram (ALLOWED)
  --> Blocked from WhatsApp external contact (BLOCKED)

Backward Trace

प्रश्न: "10:24 UTC ला पाठवलेल्या message ला कोणत्या sources contribute केले?"

Backward trace output पासून सुरू होतो आणि output ला influence केलेल्या प्रत्येक data source ओळखण्यासाठी lineage chain मधून walk back करतो. Response मध्ये classified data included होता का हे समजण्यासाठी हे essential आहे.

Output: Message sent to Telegram at 10:24:00Z
  --> session: sess_456
  --> lineage sources:
      --> lin_789xyz: Salesforce opportunity (CONFIDENTIAL)
      --> lin_790xyz: Salesforce opportunity (CONFIDENTIAL)
      --> lin_791xyz: Salesforce opportunity (CONFIDENTIAL)
      --> lin_792xyz: Weather API (PUBLIC)

Classification Justification

प्रश्न: "हा data CONFIDENTIAL का marked आहे?"

Classification justification classification level assign केलेल्या rule किंवा policy ला trace back करतो:

Data: Pipeline summary (lin_789xyz)
Classification: CONFIDENTIAL
Reason: source_system_default
  --> Salesforce integration default classification: CONFIDENTIAL
  --> Configured by: admin_001 at 2025-01-10T08:00:00Z
  --> Policy rule: "All Salesforce data classified as CONFIDENTIAL"

Compliance Export

Legal, regulatory, किंवा internal review साठी, Triggerfish कोणत्याही data element किंवा time range साठी full chain of custody export करू शकतो:

Export request:
  --> Time range: 2025-01-29T00:00:00Z to 2025-01-29T23:59:59Z
  --> Scope: All sessions for user_456
  --> Format: JSON

Export includes:
  --> All audit records in the time range
  --> All lineage records referenced by audit records
  --> All session state transitions
  --> All policy decisions (ALLOW, BLOCK, REDACT)
  --> All taint changes
  --> All delegation chain records

Compliance exports structured JSON files आहेत जे SIEM systems, compliance

dashboards, किंवा legal review tools द्वारे ingested केले जाऊ शकतात. Export format stable आणि versioned आहे. :::

Data Lineage

Audit logging Triggerfish च्या data lineage system सह conjunction मध्ये काम करते. Triggerfish द्वारे processed प्रत्येक data element provenance metadata carry करतो:

json

{
  "lineage_id": "lin_789xyz",
  "content_hash": "sha256:a1b2c3d4...",
  "origin": {
    "source_type": "integration",
    "source_name": "salesforce",
    "record_id": "opp_00123ABC",
    "record_type": "Opportunity",
    "accessed_at": "2025-01-29T10:23:45Z",
    "accessed_by": "user_456",
    "access_method": "plugin_query"
  },
  "classification": {
    "level": "CONFIDENTIAL",
    "reason": "source_system_default",
    "assigned_at": "2025-01-29T10:23:45Z",
    "can_be_downgraded": false
  },
  "transformations": [
    {
      "type": "extraction",
      "description": "Selected fields: name, amount, stage",
      "timestamp": "2025-01-29T10:23:46Z",
      "agent_id": "agent_123"
    },
    {
      "type": "summarization",
      "description": "LLM summarized 3 records into pipeline overview",
      "timestamp": "2025-01-29T10:23:47Z",
      "input_lineage_ids": ["lin_789xyz", "lin_790xyz", "lin_791xyz"],
      "agent_id": "agent_123"
    }
  ],
  "current_location": {
    "session_id": "sess_456",
    "context_position": "assistant_response_3"
  }
}

Lineage records POST_TOOL_RESPONSE वर (data system मध्ये enter झाल्यावर) created होतात आणि data transformed होताना updated होतात. Aggregated data max(input classifications) inherit करतो -- कोणताही input CONFIDENTIAL असल्यास, output कमीत कमी CONFIDENTIAL आहे.

Event	Lineage Action
Integration मधून Data read	Origin सह lineage record create करा
LLM द्वारे Data transformed	Transformation append करा, input lineages link करा
Multiple sources मधून Data aggregated	Lineage merge करा, classification = max(inputs)
Channel ला Data sent	Destination record करा, classification verify करा
Session reset	Lineage records archive करा, context मधून clear करा

Storage आणि Retention

Audit logs audit: namespace खाली StorageProvider abstraction द्वारे persisted आहेत. Lineage records lineage: namespace खाली stored आहेत.

Data Type	Namespace	Default Retention
Audit logs	`audit:`	1 year
Lineage records	`lineage:`	90 days
Session state	`sessions:`	30 days
Taint history	`taint:`	Session retention शी matches

SECURITY Retention periods configurable आहेत, पण audit logs compliance

requirements (SOC 2, GDPR, HIPAA) support करण्यासाठी default 1 year ला आहेत. तुमच्या organization च्या regulatory requirement च्या खाली retention period reduce करणे administrator ची responsibility आहे. :::

Storage Backends

Tier	Backend	Details
Personal	SQLite	`~/.triggerfish/data/triggerfish.db` वर WAL-mode database. Audit records सर्व Triggerfish state सारख्याच database मध्ये structured JSON म्हणून stored आहेत.
Enterprise	Pluggable	Enterprise backends (Postgres, S3, इ.) `StorageProvider` interface द्वारे वापरले जाऊ शकतात. हे existing log aggregation infrastructure सह integration allow करते.

Immutability आणि Integrity

Audit records append-only आहेत. एकदा written झाल्यावर, system च्या कोणत्याही component द्वारे -- LLM, agent, किंवा plugins सह -- ते modify किंवा delete केले जाऊ शकत नाहीत. Deletion फक्त retention policy expiration द्वारे होते.

प्रत्येक audit record एक content hash include करतो जो integrity verify करण्यासाठी वापरला जाऊ शकतो. Records compliance review साठी exported असल्यास, tampering detect करण्यासाठी stored records विरुद्ध hashes validate केले जाऊ शकतात.

Enterprise Compliance Features

Enterprise deployments audit logging extend करू शकतात:

Feature	वर्णन
Legal hold	Specified users, sessions, किंवा time ranges साठी retention-based deletion suspend करा
SIEM integration	Real time मध्ये Splunk, Datadog, किंवा इतर SIEM systems ला audit events stream करा
Compliance dashboards	Policy decisions, blocked actions, आणि taint patterns चे Visual overview
Scheduled exports	Regulatory review साठी Automatic periodic exports
Alert rules	Specific audit patterns occur झाल्यावर notifications trigger करा (उदा. repeated blocked write-downs)

Security-First Design -- security architecture चे overview
No Write-Down Rule -- classification flow rule ज्याचे enforcement logged आहे
Identity & Auth -- identity decisions कसे recorded होतात
Agent Delegation -- delegation chains audit records मध्ये कसे appear होतात
Secrets Management -- credential access कसे logged आहे

Audit & Compliance ​

काय Recorded होते ​

Audit Record Examples ​

Allowed Output ​

Blocked Write-Down ​

Taint Escalation सह Tool Call ​

Agent Delegation Blocked ​

Audit Trace Capabilities ​

Forward Trace ​

Backward Trace ​

Classification Justification ​

Compliance Export ​

Data Lineage ​

Storage आणि Retention ​

Storage Backends ​

Immutability आणि Integrity ​

Enterprise Compliance Features ​

Related Pages ​

Audit & Compliance

काय Recorded होते

Audit Record Examples

Allowed Output

Blocked Write-Down

Taint Escalation सह Tool Call

Agent Delegation Blocked

Audit Trace Capabilities

Forward Trace

Backward Trace

Classification Justification

Compliance Export

Data Lineage

Storage आणि Retention

Storage Backends

Immutability आणि Integrity

Enterprise Compliance Features

Related Pages