What is source-sink analysis for AI agents?
Source-sink analysis, applied to AI agents, maps the sources through which untrusted content can influence an agent and the sinks, such as sending data or following links, that become dangerous if the agent is misled.
Source-sink analysis is an established security engineering technique: track where untrusted data enters a program (the source) and where it could cause harm (the sink), then check every path between them. Static analysis tools call the tracking part taint tracking, which follows untrusted data even as it is transformed.
OpenAI applied the idea to agents in its March 2026 post on designing agents to resist prompt injection. The post treats prompt injection as social engineering aimed at the model and, for ChatGPT, combines that view with source-sink analysis. In OpenAI's usage, a sink is a capability that becomes hazardous once the model has been misled, such as sending information to a third party, following a link or using a tool.
Sources and sinks in an agent
Sources
Web pages, emails, shared documents, support tickets, file contents, tool and MCP server outputs, and messages from other agents: anything the agent reads that a third party could have written.
Sinks
Sending email or messages, HTTP requests and link fetches (a URL can carry data out in its query string), writes to shared systems, payments and code execution.
Paths
The route from a source to a sink through the agent's context. Because a model reads everything in one context, any source can influence any sink unless the design separates them.
Why it beats input filtering alone
Classifiers that scan inputs for injected instructions catch known patterns and miss new ones. OpenAI's position is that a successful manipulation should still have limited impact, which means removing or guarding paths rather than relying on detection.
OpenAI describes one sink-side mitigation in ChatGPT, called Safe Url, which detects when the assistant is about to send information from the conversation to a third party. The defence sits at the sink, where it works whichever source the attack came from.
Related framings
Lethal trifecta
Simon Willison (June 2025): an agent with private data, exposure to untrusted content and a way to communicate externally is at serious risk. Remove one leg.
Agents Rule of Two
Meta (October 2025): within one session, an agent should have at most two of three properties: processing untrustworthy input, accessing sensitive systems or private data, and changing state or communicating externally. If it needs all three, it needs supervision or the task should be split across sessions.
CaMeL
Google DeepMind's 2025 design tracks data flow and capabilities in a custom interpreter, so the source-sink check is enforced by code at every tool call. See the dual LLM pattern.
Doing the analysis
List the sources
Every input path, including tool results and retrieved documents, not only the user's message.
List the sinks
Every tool that sends, writes, pays, executes or fetches.
Draw the paths
For each sink, which sources can influence its arguments within the same context.
Cut or guard each path
Split reading and acting into separate sessions, allowlist destinations, require approval, or remove the sink from that agent altogether.
How Motiontic's engineers apply it
Motiontic's engineers put a source-sink map in the design document of every agent that reads outside content, and review it with the client's security team. Each path from an untrusted source to a high-impact sink is either removed, by splitting reading and acting into separate steps, or guarded by an allowlist or an approval, and injection tests target every path that remains.
Three framings of the same problem
| Source-sink analysis | Lethal trifecta | Agents Rule of Two | |
|---|---|---|---|
| From | Security engineering, applied to agents by OpenAI (2026) | Simon Willison (2025) | Meta (2025) |
| Unit of analysis | Paths from sources to sinks | Three capabilities in one agent | Three properties in one session |
| Rule | Limit what a successful manipulation can reach | Remove one of the three | Hold at most two, or add supervision |
| Granularity | Per path and per sink | Per agent | Per session |
Common questions
What are sources and sinks in prompt injection?
A source is any way untrusted content can reach the agent, such as a web page, email or tool output. A sink is a capability that does harm if the agent is misled, such as sending data out, following a link or calling a tool.
How is source-sink analysis different from the lethal trifecta?
The lethal trifecta judges an agent as a whole by three capabilities. Source-sink analysis works path by path, which lets a design keep a risky capability while guarding the specific routes an attack could take to it.
Does source-sink analysis prevent prompt injection?
No. It assumes some injections will succeed and limits what they can reach. OpenAI and the other labs describe prompt injection as reduced, not solved.
What is taint tracking?
Taint tracking extends data flow analysis to follow untrusted data even when its value changes, so anything derived from a tainted input is treated as tainted too. CaMeL applies a similar idea to agent tool calls.
Related
Prompt injection
Untrusted text that hijacks a model's instructions, and how agents are built to survive it.
Dual LLM pattern (privileged and quarantined LLMs, CaMeL)
Keep the model that acts away from the text an attacker controls.
Blast radius (AI agents)
The worst an agent could do, measured by what it can reach.
Agent sandbox
Hard walls around where an agent can read, write and connect.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- OpenAI, Designing AI agents to resist prompt injection (2026-03-11)
- Simon Willison, The lethal trifecta for AI agents (2025-06-16)
- Meta, Agents Rule of Two: a practical approach to AI agent security (2025-10-31)
- Debenedetti et al., Defeating prompt injections by design (CaMeL, arXiv, 2025-03-24)
- GitHub CodeQL, About data flow analysis
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.