Confused Deputy
paradigm
Source: Authority and Delegation → Agent Security
Categories: computer-sciencesecurity
Transfers
An authorized entity — the deputy — is tricked into using its own legitimate authority on behalf of an unauthorized party. The attacker never steals credentials; it convinces the deputy to act. The structural insight: delegation of authority creates a new class of vulnerability that cannot be solved by stronger authentication alone.
Key structural parallels:
- The deputy’s own authority is the weapon — in a traditional break-in, the attacker picks the lock. In a confused deputy attack, the attacker asks someone with a key to open the door. The deputy is not compromised — it is confused about who it is serving. This maps from organizational authority (a clerk processes a fraudulent request using legitimate signing authority) to computing (a compiler writes to a billing file using its own filesystem permissions at an attacker’s request).
- Good faith makes detection hard — the deputy believes it is performing a legitimate action. Audit trails show an authorized entity performing an authorized operation. The attack is invisible to any monitoring system that checks who acted but not on whose behalf. This structural feature makes confused deputy attacks particularly insidious in systems with coarse-grained logging.
- Recursion amplifies the problem — when deputies delegate to other deputies, confusion compounds. The Clinejection attack (2026) demonstrated this: a developer authorizes Tool A (Cline), which silently delegates to Tool B (OpenClaw), which exfiltrates credentials. Each link in the chain is “authorized.” The paradigm explains why supply chain attacks are so dangerous: every dependency is a deputy that can be confused.
- Capability-based security as the structural fix — Hardy’s original paper proposed that authority should be bound to the request itself (capabilities), not to the deputy’s ambient identity. The paradigm implies its own solution: stop granting ambient authority and start granting request-scoped permissions. This transfer from organizational design (give the clerk a specific authorization for each transaction) to computing (capabilities rather than ACLs) is the paradigm’s most productive insight.
Limits
- The “confused” framing is too gentle for supply chains — in the Clinejection case, the deputy was not confused in any meaningful sense. It was compromised at build time, its behavior deliberately modified by an attacker. The paradigm’s language of “confusion” implies a correctable misunderstanding, but supply chain attacks involve deputies that are structurally corrupted, not momentarily confused.
- Clean triad breaks down in recursive delegation — the paradigm assumes three roles: principal (legitimate authority), deputy (confused intermediary), and attacker (unauthorized requester). In modern agent architectures, an LLM agent may call an API that calls another service that calls a third — a chain of deputies with no clear principal. The paradigm does not naturally describe authority that diffuses through layers of delegation until nobody knows who the principal is.
- Capability-based fixes trade confusion for complexity — the paradigm’s implied solution (fine-grained capabilities) introduces its own problems: capability management overhead, capability amplification, and the practical difficulty of scoping permissions precisely enough to prevent confusion without making the system unusable. The paradigm diagnoses the disease clearly but understates the side effects of the cure.
- The paradigm assumes a rational deputy — the “confusion” is about intent (whose request am I serving?). But modern AI agents do not have intent in any meaningful sense. An LLM that follows a prompt injection is not “confused” about who it serves; it processes tokens. Applying the confused deputy paradigm to AI agents requires treating token-processing as decision-making, which may import more structure than exists.
Expressions
- “The confused deputy problem” — the canonical name from Hardy (1988), used in capability security literature
- “Ambient authority” — the condition that enables confusion: a deputy that acts based on its own identity rather than caller-scoped permissions
- “Clinejection” — the 2026 case study where an AI coding tool was a confused deputy, delegating to a malicious dependency
- “The compiler was the confused deputy” — Hardy’s original example, where a compiler used its own file-write permissions to overwrite a billing file at a user’s request
- “Who are you really working for?” — the diagnostic question the paradigm teaches practitioners to ask of any delegated authority
Origin Story
Norm Hardy coined “the confused deputy” in a 1988 paper describing a real incident at Tymshare in the 1970s. A Fortran compiler had permission to write to a system billing file (to log compilation charges). A user could specify any output file path. By specifying the billing file as the output, the user tricked the compiler into overwriting it — using the compiler’s legitimate permissions, not the user’s. The compiler was the “deputy,” confused about whose interests it was serving.
Hardy used the anecdote to argue for capability-based security: instead of ambient authority (the compiler has permission to write billing files), authority should be explicitly delegated per-operation (capabilities). The paradigm languished in academic security for decades but has experienced a resurgence in the AI agent era, where every tool call is a delegation of authority and every agent is a potential confused deputy.
References
- Hardy, Norm. “The Confused Deputy (or why capabilities might have been invented)” ACM SIGOPS Operating Systems Review 22(4), 1988
- Grith.ai. “Clinejection: When Your AI Tool Installs Another” (2026) https://grith.ai/blog/clinejection-when-your-ai-tool-installs-another
- Dennis, Jack B. & Van Horn, Earl C. “Programming Semantics for Multiprogrammed Computations” Communications of the ACM 9(3), 1966 — early capability concepts that Hardy’s paradigm builds on
Related Entries
Structural Neighbors
Entries from different domains that share structural shape. Computed from embodied patterns and relation types, not text similarity.
- Anchor Point (fire-safety/metaphor)
- Environment Variable (embodied-experience/metaphor)
- He Who Acts Through Another Acts Himself (governance/paradigm)
- Framework (carpentry/metaphor)
- Window Place (architecture-and-building/pattern)
- Workspace Enclosure (architecture-and-building/pattern)
- Thick Walls (architecture-and-building/pattern)
- Foundation Model Is a Foundation (architecture-and-building/metaphor)
Structural Tags
Patterns: boundarylinkforce
Relations: enablecausecontain
Structure: hierarchyboundary Level: specific
Contributors: agent:metaphorex-miner