Audience: Security architects deploying agentic AI systems, AI platform engineers implementing multi-agent orchestration, federal agencies evaluating agent governance frameworks, and CISOs assessing risk from autonomous AI agents.
Emerging framework. Not a final standard. The CAISI AI Agent Standards Initiative launched February 17, 2026, organized around three pillars: agent security, interoperability, and identity. The NCCoE published a complementary concept paper (February 5, 2026) covering identification, authorization, access delegation, and logging. NIST red team results show an 81% task-hijacking success rate against unprotected agents (vs. 11% baseline). These documents form a coherent threat model but have not been consolidated into a single standard. This guide maps the emerging requirements to SWT3 witness procedures that are available today.
Contents
1. The Agentic AI Threat Landscape 2. NIST Initiative Structure 3. Obligation-to-Procedure Mapping 4. SWT3 Procedure Cards 5. Multi-Agent Chain Evidence 6. Quick Reference 7. Quick Start 8. References1. The Agentic AI Threat Landscape
Autonomous AI agents that call tools, delegate tasks to sub-agents, and make sequential decisions without human intervention represent a fundamentally different security surface than traditional AI inference. An agent is not just a model producing outputs -- it is a model acting on those outputs, with real-world consequences.
NIST's own red team research found that novel attack techniques targeting AI agents achieved an 81% task-hijacking success rate, compared to 11% for the strongest known baseline attacks. The gap between existing defenses and emerging threats is not incremental. It is structural.
Core risks:
- Identity spoofing: Agents impersonating other agents or humans to gain elevated access
- Privilege escalation through delegation: A low-privilege agent delegating to a higher-privilege tool without authorization checks
- Unaudited tool calls: Agents invoking external tools (APIs, databases, file systems) without witnessing what was called, what was returned, and what was done with the result
- Chain-of-custody gaps: Multi-step agent workflows where intermediate decisions are not recorded, making forensic reconstruction impossible after an incident
- Prompt injection via tool results: Adversarial content in tool responses redirecting agent behavior
Traditional security frameworks were designed for request-response architectures. Agentic systems require identity, delegation, and audit infrastructure purpose-built for autonomous multi-step workflows. This is the gap NIST is working to close.
2. NIST Initiative Structure
| Pillar | Focus Areas | Status |
|---|---|---|
| Agent Security | Attack surfaces, prompt injection, tool poisoning, sandbox escape | CAISI research active |
| Interoperability | Cross-platform agent communication, protocol standards, NSF-funded open-source | Community development |
| Identity | Agent identification, credential management, delegation tracking, access logging | NCCoE concept paper published |
NCCoE Concept Paper: Four Focus Areas
| Focus Area | Description | SWT3 Coverage |
|---|---|---|
| Identification | Every agent must have a verifiable, unique identity | AI-ID.1 |
| Authorization | Agent actions must be scoped to authorized capabilities | AI-ACC.1 |
| Access Delegation | When agents delegate to sub-agents, the delegation chain must be recorded | AI-DEL.1 |
| Logging | All agent actions must produce structured, retrievable event logs | AI-LOG.1 |
SWT3 is the only open protocol that provides cryptographic witness infrastructure across all four NCCoE focus areas. Each agent action produces an immutable Witness Anchor with a cross-language fingerprint formula, creating the audit trail NIST envisions but has not yet standardized.
3. Obligation-to-Procedure Mapping
| NIST Requirement | Evidence Needed | SWT3 Procedure |
|---|---|---|
| Agent identification | Unique agent_id per autonomous entity | AI-ID.1 |
| Action authorization | Scoped access records per tool call | AI-ACC.1 |
| Delegation tracking | Parent-child delegation trees | AI-DEL.1 |
| Tool call witnessing | Input/output records per external tool invocation | AI-TOOL.1 |
| Structured logging | Retrievable event logs with timestamps | AI-LOG.1 |
| Chain of custody | Linked anchor sequences across multi-step workflows | AI-CHAIN.1 |
| Security boundary | Isolation evidence between agents | AI-SEC.1 |
| Audit readiness | Tamper-evident operational records | AI-AUDIT.1 |
4. SWT3 Procedure Cards
Agent Identity
NIST context: The NCCoE concept paper identifies agent identification as the first focus area. Every autonomous entity must have a verifiable, unique identity that persists across sessions and can be authenticated by other agents and oversight systems.
SWT3: Every witness call includes an agent_id field. The SDK's WitnessClient(agent_id="agent-name") binds a verifiable identity to every action the agent takes. Identity is recorded in every Witness Anchor, creating an immutable record of which agent did what.
AI-ID.1 is the identity root for all other agent procedures. Verify that agent_id values are unique, consistent, and not shared across unrelated agents. Multiple distinct agents sharing an agent_id is an identity integrity finding.
Access Control
NIST context: Agent actions must be scoped to authorized capabilities. An agent authorized to read a database should not be able to write to it. An agent with access to one API should not be able to call arbitrary others.
SWT3: wrap_access() / wrapAccess() witnesses access control decisions. Records what resource was accessed, what permission was evaluated, and whether access was granted or denied. Creates evidence of least-privilege enforcement.
AI-ACC.1 anchors prove access was scoped. Look for access_granted=true records without corresponding authorization_id values -- this indicates access without pre-authorization, which NIST identifies as a privilege escalation vector.
Delegation Trees
NIST context: When agents delegate tasks to sub-agents, the delegation chain must be recorded. Untracked delegation is the primary mechanism for privilege escalation in multi-agent systems -- a low-privilege orchestrator delegating to a high-privilege tool agent without oversight.
SWT3: Delegation events produce anchors linking parent and child agents via cycle_id. The audit portal renders delegation trees visually, showing the complete chain from the initiating agent through every sub-agent. swt3 reconstruct --cycle rebuilds the full delegation history.
AI-DEL.1 anchors are critical for multi-agent systems. Verify that every sub-agent action has a parent delegation anchor. Orphaned agent actions (no delegation parent) in a multi-agent system indicate untracked delegation -- a serious finding.
Tool Call Witnessing
NIST context: Agents invoke external tools -- APIs, databases, file systems, code interpreters. Each tool call is an attack surface. NIST red team research demonstrates that adversarial content in tool responses can hijack agent behavior at an 81% success rate.
SWT3: wrap_tool() / wrapTool() witnesses every tool invocation: what tool was called, what arguments were passed, what was returned, and whether the result was used. Creates a forensic record of the agent's interaction with external systems.
AI-TOOL.1 anchors prove tool calls were witnessed. For NIST compliance, verify that tool_name, tool_input, and tool_output are all recorded. Tool calls without output records suggest the witnessing infrastructure was bypassed or the tool response was consumed without recording.
Logging Pipeline
NIST context: All agent actions must produce structured, retrievable event logs. The NCCoE concept paper lists logging as the fourth focus area -- foundational infrastructure without which the other three areas cannot be verified.
SWT3: Records logging pipeline health -- completeness, delivery confirmation, and structured event capture. The OTel exporter integrates with existing observability infrastructure (Datadog, Splunk, Grafana) so agent logs flow into the same pipeline as application telemetry.
AI-LOG.1 anchors prove the logging infrastructure was operational. Gaps in AI-LOG.1 during periods with active agent operations indicate logging failures -- the most fundamental infrastructure finding.
Chain of Custody
NIST context: Multi-step agent workflows must maintain chain of custody -- every intermediate decision linked to its predecessors and successors. Without chain integrity, forensic reconstruction after an incident is impossible.
SWT3: Lifecycle chains link multiple anchors via shared cycle_id. Each anchor in the chain records its position in the sequence, creating a verifiable timeline of the entire workflow. swt3 reconstruct --chain rebuilds the chain for forensic analysis.
AI-CHAIN.1 is the connective tissue. Verify that multi-step operations share a cycle_id and that the chain is complete (no missing sequence numbers). Broken chains indicate lost evidence.
Security Boundary
NIST context: Agents must operate within defined security boundaries. Sandbox escape -- where an agent breaks out of its intended execution environment -- is a critical attack vector identified in NIST research.
SWT3: Records security boundary status -- what isolation was in place, whether containment was maintained, and whether any boundary violations were detected. Creates evidence that agents operated within their authorized scope.
AI-SEC.1 anchors demonstrate security boundary enforcement. For containerized agents, verify that boundary_type reflects the actual isolation mechanism (container, VM, process). Self-reported boundary claims without infrastructure verification are insufficient.
Audit Trail
NIST context: Tamper-evident audit trails are the ultimate deliverable for agent governance. NIST's agent standards initiative assumes organizations can reconstruct agent behavior for any historical period. Without immutable records, all other controls are unverifiable.
SWT3: Every Witness Anchor is an audit trail entry. Cryptographic fingerprints (SHA-256) ensure immutability. The append-only ledger with daily Merkle rollups provides both granular and aggregate integrity verification. Forensic timeline reconstruction via swt3 reconstruct enables full operational replay.
AI-AUDIT.1 is the capstone. Verify anchor continuity -- gaps in the audit trail during active agent operations are the most significant finding. Use swt3 verify --enclave for integrity verification across all anchors.
5. Multi-Agent Chain Evidence
Production agentic systems rarely involve a single agent. An orchestrator delegates to specialist agents, which invoke tools, which return results that feed into downstream decisions. SWT3's lifecycle chain infrastructure links all of these actions into a single, reconstructable timeline.
Forensic Timeline Reconstruction
The swt3 reconstruct CLI command rebuilds multi-agent workflows from anchor data. Query by cycle_id (complete workflow), agent_id (single agent's history), or time range (incident window). Output includes delegation trees, drift markers, override flags, and clearing level badges.
# Reconstruct a complete multi-agent workflow
swt3 reconstruct --cycle cyc_abc123
# View all actions by a specific agent
swt3 reconstruct --agent agent-orchestrator-1
# Export as self-contained HTML report
swt3 reconstruct --cycle cyc_abc123 --html
# Query via API
curl -H "Authorization: Bearer axm_live_..." \
"https://sovereign.tenova.io/api/v1/reconstruct?cycle_id=cyc_abc123"Delegation Tree Visualization
The audit portal renders delegation trees visually, showing the complete chain from the initiating agent through every sub-agent. Each node displays the agent_id, procedure, verdict, and timestamp. Drift and override markers are highlighted for immediate attention. Assessors can expand any node to view the full anchor detail.
6. Quick Reference
| Security Architect Question | Where to Look |
|---|---|
| Does every agent have a unique, verifiable identity? | AI-ID.1 anchors. Verify agent_id uniqueness across the deployment. Shared IDs = identity integrity gap. |
| Are tool calls witnessed with input and output? | AI-TOOL.1 anchors with tool_name, tool_input, tool_output fields. Missing output records indicate bypassed witnessing. |
| Can we reconstruct a multi-agent workflow after an incident? | swt3 reconstruct --cycle <cycle_id>. Complete chain = all sequence numbers present. Broken chain = lost evidence. |
| Are delegation chains tracked from orchestrator to sub-agent? | AI-DEL.1 anchors with parent-child agent relationships. Orphaned agent actions = untracked delegation. |
| Is the logging pipeline monitored for gaps? | AI-LOG.1 anchors showing continuous pipeline health. Gaps during active operations = logging failure. |
| Are security boundaries enforced and verified? | AI-SEC.1 anchors with boundary_type and containment_status. Self-reported claims without infrastructure verification are insufficient. |
7. Quick Start
# Install the SDK
pip install swt3-ai
from swt3_ai import WitnessClient
client = WitnessClient(
tenant_id="your-tenant-id",
api_key="axm_live_...",
agent_id="agent-orchestrator-1"
)
# Witness a tool call
client.wrap_tool(
tool_name="database_query",
tool_input={"query": "SELECT * FROM users WHERE active=true"},
tool_output={"rows": 42}
)
# Witness delegation to a sub-agent
client.witness_delegation(
child_agent_id="agent-analyst-2",
delegated_task="analyze_user_activity",
cycle_id="cyc_abc123"
)
# Witness access control
client.wrap_access(
resource="user_database",
action="read",
granted=True,
authorization_id="auth_xyz789"
)
# Run the demo to see it in action
python -m swt3_ai.demoSWT3 SDK Documentation | Create a free account
8. References
- NIST CAISI AI Agent Standards Initiative
- NCCoE Concept Paper: Agentic AI Identity and Authorization
- NIST Blog: Insights into AI Agent Security from a Large-Scale Red-Teaming Competition
- CSA Research Note: NIST AI Agent Security
- Singapore IMDA Agentic AI Crosswalk (SWT3 Protocol)
- OpenAI Agentic Governance Crosswalk (SWT3 Protocol)
- Multi-Agent Chain Audit Trail Guide (SWT3 Protocol)
- SWT3 SDK Documentation
- Create a free account