The Agentic AI Autonomy Levels and Control Framework is a white paper from the Cloud Security Alliance AI Safety Initiative, published on the CSA Labs site on 18 March 2026 as Version 2.0. It sets out, in its own words, “a comprehensive framework for defining, implementing, and governing AI autonomy levels,” with the aim of “ensuring that autonomy is granted deliberately rather than by default.” The paper states that Version 2.0 revises an initial publication of 29 January 2026, drawing on ten security incidents documented in the fifty days between the two.
The paper describes its own status plainly. Its cover and footers read “Unofficial AI-assisted Research,” and its notice page states: “This document was generated with AI assistance and has not undergone official CSA review and approval processes.” This page treats it accordingly: as published research from a CSA initiative, mapped as it stands, and not as ratified CSA guidance.
This page is a mapping, not a review. It maps the framework’s principles, autonomy levels, control categories and v2.0 extensions onto this doctrine’s constructs, records three places where the two differ by design, and notes what the doctrine’s assessment instruments provide alongside the framework. The framework does not cite this doctrine, and this mapping does not suggest otherwise.
What was read. The CSA Labs landing page for the framework was read on 3 November 2026. The PDF (Version 2.0 | March 2026, dated 18 March 2026 on its title page, 73 pages) was read in full on 3 November 2026, including its appendices. Not read: the two CSA companion documents the paper relies on for its incident evidence, the Post-Incident Update Assessment and The Cost of Unchecked Autonomy; Version 1.0 of the framework; any other source the paper cites; and the CSA frameworks it integrates with in Appendices C to E (CBRA, the AI Controls Matrix and MAESTRO), which are described here only as the paper describes them. This page makes no claim about the incidents themselves. Where this page and the framework disagree about the framework, the framework is right.
The five principles
The framework names five principles that “underpin the framework’s approach to autonomy governance.” Each has a direct counterpart in this doctrine. Left column is CSA’s wording, verbatim.
| CSA principle | Where it lands in this doctrine |
|---|---|
| “autonomy should be explicitly defined and justified rather than granted implicitly or by default” | Least Agency, the third of the three non-negotiables: agents operate with “no greater decision scope or authority than their approved missions require,” and Least Agency “includes the zero case: the mission owner first decides whether autonomous decision-making is warranted.” The approved mission itself is protected under R03 |
| “higher autonomy requires correspondingly stronger controls to manage the increased risk” | The doctrine’s effect classes: recovery capability, sensitivity, affected parties, magnitude, time to harm and context inform how each effect is classified, and “Expanding a declared mission boundary shall not silently lower the required protection” |
| “human oversight should be proportional to the risk posed by autonomous actions” | Law 5, Humans Retain the Right to Intervene; profiles state which effects require action-specific human approval and which can proceed under an already approved bounded class |
| “autonomy boundaries must be technically enforced rather than relying solely on policy” | The first two non-negotiables: governance operates in the execution path, and enforcement covers the declared boundary “rather than rely on a fragment that the agent can bypass”; Law 2, Enforce at Runtime; R06, effects bounded by external controls |
| “continuous monitoring is essential at all autonomy levels to detect anomalies and ensure appropriate operation” | Plane 4, Human Oversight, Audit & Traceability; R16 and R17 for protected decision evidence and auditable actions |
The six autonomy levels
The framework’s taxonomy classifies an agent from Level 0 to Level 5. This doctrine does not define autonomy levels of its own (see the next-but-one section), so the right column records the doctrine construct that governs what each level permits, not an equivalent level. Left column is each level’s name from the framework, verbatim, with its description summarized.
| CSA level | Where it lands in this doctrine |
|---|---|
| Level 0: No Autonomy (Human Execution). The AI provides information; humans make all decisions and execute all actions. | The zero case of Least Agency: the mission owner decides that autonomous decision-making is not warranted |
| Level 1: Assisted (Human Decision + AI Execution). Each action requires explicit human approval. The framework calls it “the recommended starting point for organizations new to agentic AI.” | Effects classified as requiring additional authorization; R07: required human authorization binds the approved action, and “an agent-authored narrative alone shall not establish what was approved” |
| Level 2: Supervised (Human Approval + Batch Execution). Humans approve a plan; execution within it proceeds automatically. | R07: authority “may cover a single action or recurring work within declared bounds,” and protected enforcement binds the evaluated effect to what executes; pause and cancel at R15 |
| Level 3: Conditional (AI Decision within Boundaries). The AI acts within defined boundaries and escalates when they are exceeded. The framework calls it “a significant governance threshold.” | Effects permitted within standing bounds; R06, where external controls “deny unknown, unclassified or unenforceable effects”; Plane 2, Execution & Tool Governance |
| Level 4: High Autonomy (Minimal Supervision). Broad autonomous operation, with human oversight based on monitoring rather than approval. | Law 5, which requires a profile to identify “measurable intervention behavior”; R10 for effective revocation; R15 for interruption with accountable continuity |
| Level 5: Full Autonomy (Self-Directed). Includes goal-setting and self-modification. The framework says it “is not recommended for current enterprise deployments.” | The doctrine’s definition of a mission: “an approved bounded operational assignment, not an organizational slogan or an agent’s self-description”; Law 2: standing authorization “does not remove current automated checks or create permanent trust” |
The six control categories
The framework organizes its required controls into six categories and specifies the controls in each category for each level. Left column is CSA’s category name, with its purpose as the paper states it.
| CSA control category | Where it lands in this doctrine |
|---|---|
| Authorization. Who or what can grant autonomy to an AI system. | Plane 1, Agent Identity & Lifecycle, which binds approved missions and grants; R02 and R03 |
| Boundary. Technical limits on what actions an agent can take. | Plane 2; R06; R11 for cumulative and sequence-sensitive limits |
| Oversight. Monitoring and intervention mechanisms. | Law 5; Plane 4; Framework Pillar 5, Human Oversight & Intervention |
| Accountability. Logging and attribution of actions. | Law 1, Agents Are Identities, Not Tools; R16, under which “Agent narration alone shall not attest that an effect occurred”; R17 |
| Recovery. Reversal and remediation when issues occur. | R15, which distinguishes pause, cancel, revoke, quarantine, handoff and recovery; R10, covering queued, in-flight, delegated and irreversible effects |
| Governance. Policy and process requirements. | Plane 3, Policy & Compliance Engine; R01, the declared deployment boundary; R18 for exceptional operation |
The v2.0 extensions
Version 2.0 adds several sections the paper marks as new. The ones below have a direct counterpart in this doctrine. Left column quotes the framework, verbatim.
| CSA v2.0 extension | Where it lands in this doctrine |
|---|---|
| Autonomy Escalation Prevention (Section 7). Four principles: self-modification prohibition (“An agent must not be able to modify its own autonomy level”), self-disabling prohibition (“An agent must not be able to disable its own oversight controls”), self-expansion prohibition (“An agent must not be able to expand its own scope through its tool set”), and architectural separation (“Autonomy level configuration must be enforced at an architectural layer the agent’s execution context cannot reach”). | The Architecture’s enforcement and decision functions: “The agent shall not be able to replace or bypass the functions that govern it,” and the claimed separation “must survive the actor and paths within the assessed boundary.” R06: “The acting agent shall not be able to alter its governing policy or controls through direct or indirect paths.” The framework draws the same analogy to control-plane and data-plane separation that motivates Planes 2 and 3. |
| Supply Chain Trust Delegation (Section 6). “any operation that expands the agent’s capability set is an autonomy-expanding operation that requires controls proportional to the autonomy the installed component grants, not just the autonomy the agent currently holds.” | Law 1: an agent’s “relevant model and tools” shall be distinguishable; R01: the assurance profile names “the model, tool and integration bindings on which its claims depend”; R24: changes to models, tools or integrations trigger impact review and proportionate retesting |
| Delegation Scope Narrowing (Section 8.3). “Each delegation hop in a multi-agent system should reduce the available autonomy, not maintain or expand it.” | Law 4, Trust Does Not Travel: delegation “shall not amplify what the delegator is authorized to grant”; R08; Plane 5, Multi-Agent Trust & Delegation. See the next section for one difference in mechanism. |
| Composite Autonomy Level (Section 8.2). “Organizations must assess and authorize the composite autonomy level of their multi-agent deployments, not merely the individual levels of constituent agents.” | Law 3, Governance Must Span Systems, whose declared boundary includes delegated work and downstream effects; R11: cumulative limits “across agents, delegations, retries, destinations and time windows,” with shared counters that prevent “handoff-based resets” |
| Reversibility Refinement (Section 2.1.1). “The control requirements in this framework are driven by consequence reversibility, not action reversibility alone.” | The doctrine’s effect classes: a read “can disclose data,” its treatment depends “on the actual data and recipient boundary, not a tool’s label,” and “Detection latency alone does not establish recoverability”; R15: resumption shall “prevent unintended repetition of irreversible effects” |
| Prompt Injection as Autonomy-Level Attack (Section 10). Prompt injection framed as “an exploit class whose effect is to escalate the agent’s operational autonomy level without authorization,” with external boundary enforcement as defense in depth. | Prompt injection is the first of the four threats retained under Threats and limits, alongside intent hijacking, cascading failure and behavioral drift; profiles also consider authority expansion. External enforcement at R06 is the doctrine’s structural answer |
| Inter-Agent Communication Monitoring (Section 8.4). “All inter-agent messages must be logged to a human-accessible audit trail in real time.” | Plane 5, which preserves grant chains and affected-output lineage; R13; R16 |
| Training-Time Autonomy Controls (Section 11). Network egress restrictions, resource hard limits, behavioral anomaly monitoring and environment isolation for agents in training with real-world connectivity. | R01: a declared boundary in which “An uncovered path shall not be represented as governed”; the Assessment Companion: “A sandbox or a simulation has its own boundary and cannot inherit a claim about production” |
Where the two differ by design
What is classified. The framework classifies the agent: each deployment is assigned a level from 0 to 5, and the level determines the required controls, with a Capability-Control Matrix setting a maximum recommended level for each capability’s risk class. This doctrine classifies effects instead. A deployment profile classifies each effect “as permitted within standing bounds, requiring additional authorization, or prohibited,” and one agent can carry effects in all three classes at once. The two are compatible. A framework level can be read as a summary of a profile’s effect classes, and the framework itself notes that a system “might have high autonomy on one dimension while being constrained on another.”
Who may authorize. The framework assigns authorization authority by level, from the business owner at Levels 0 and 1 to “CTO/CISO + Board” at Level 4, with a review cadence that rises from annual to weekly. This doctrine leaves organizational roles to the adopter. Its text states that “No universal dollar threshold or organizational level determines permission,” and the Organizational Maturity Evaluation provides that a maturity rating “shall not license an action.” What the doctrine requires is that authority be explicit, bounded and attributable, by R02, R03 and R18. An adopter can adopt the framework’s authorization hierarchy as the organizational rule its profile declares.
How delegated authority narrows. The framework requires each delegation hop to reduce autonomy, and directs enforcement infrastructure to verify that “the requested delegation scope is a subset of the delegator’s authorized scope.” This doctrine’s Law 4 permits delegation to narrow or exercise an authorized grant chain, permits federation in which “different resource authorities can grant the authority for their resources,” and notes that “An orchestrator need not accumulate every execution privilege held by its workers.” Both rule out privilege amplification down a chain. They differ on whether a worker’s authority must always be carved from its orchestrator’s, and whether a given deployment satisfies R08 is a question for assessment against a declared profile, not one this page settles.
What this doctrine provides alongside it
The framework already reaches toward assessment. It gives level-specific readiness checklists, a table of verification tests for escalation prevention (an agent attempts to modify its own boundary specification, and the expected result is “Attempt blocked; alert generated”), and a certification pathway in which “Standard certification requires third-party assessment of controls.” For deciding how much autonomy a use case warrants, and which controls to build first, it is the more detailed of the two documents.
The doctrine’s contribution sits at the point where a claim is checked. The mandatory runtime baseline of Doctrine v4 and its Assessment Companion specify what an assessor records and how a verdict is reached, for a named deployment rather than an autonomy level:
| Concern | What the doctrine specifies |
|---|---|
| Whether a passed check can be trusted | R20: each criterion has applicability conditions, evidence requirements, a method, expected outcomes, failure rules and retest triggers, and “A yes/no question is insufficient if a system could violate the property and still answer yes” |
| Telling a missing control from missing evidence | The Assessment Companion’s evidence states: a confirmed absent control fails a criterion, while unavailable evidence leaves the property not assessed and “cannot support a conformance claim” |
| What a claim covers | R23: each claim names criterion and profile versions, scope, date, evidence class, assessor and limitations, and “only a specified operated deployment can support a deployment claim” |
| Who can verify the record | R16: evidence is tamper-evident and checkable by “a verifier other than the acting agent”; “agent self-attestation alone is insufficient” |
| What happens when the governing system is unavailable | R14: ordinary effects whose required checks cannot be satisfied do not proceed, and continued operation uses only a pre-authorized bounded continuity mode, “never an implicit fail-open bypass” |
| Whether governance still lets the work get done | R22: validation measures legitimate completion and human burden alongside prohibited-effect prevention, and “a composite score shall not hide an unsafe commit” |
An autonomy-level framework and a conformance baseline answer different questions. The first says how much autonomy an agent should be granted and which controls that grant requires; the second says what evidence would show that a named deployment is operating within the authority it was granted. They are designed to be used together: a framework level and its control set can be declared in an assurance profile, and the baseline and Assessment Companion state how a third party would check it. The Cloud Security Alliance’s AARM specification is mapped separately on this site, at Governed Autonomy and AARM; this page asserts no relationship between the two CSA documents.
Primary sources
- Cloud Security Alliance AI Safety Initiative, Agentic AI Autonomy Levels and Control Framework, CSA Labs landing page (published 18 March 2026; read 3 November 2026)
- The framework, Version 2.0, as PDF (dated 18 March 2026, 73 pages; read in full 3 November 2026; the source for every framework quotation on this page)
Mapped 3 November 2026 against Version 2.0 of the framework (March 2026). When CSA revises the framework, this page will be revised to match it rather than defended.