MIT Technology Review Insights publishes Cyberhaven-sponsored report arguing data lineage is key to securing enterprise AI agents

No AI model is named. The relevant products are Cyberhaven's:

• Agentic AI Security (May 2026) has three parts: discovery of agents, GenAI apps and MCP servers, with a "Risk IQ" score; observability of data accessed, tools called and actions taken; and runtime policy controls (Cyberhaven). It shipped with an Analyst Plugin offering 40+ security skills and a standalone browser extension (IT Brief).
• Flow records agents' prompts, tool calls, file reads and responses. It integrates with Anthropic's Claude Compliance APIs and OpenAI's ChatGPT Enterprise Compliance API. It is in early access (Cyberhaven).

• Unit of control. The report's argument is that security should follow the data, not the system holding it. Content gets ingested, summarized, transformed and redistributed between agents, so labels attached to the original file stop applying (MIT TR).
• Telemetry. Cyberhaven's design captures data at the endpoint and browser and links each agent action to where the data came from and where it went next (Cyberhaven).
• Related controls. OWASP-aligned guidance suggests logging the source and timestamp of every agent memory write, so poisoned entries can be rolled back. It also suggests extending software bills of materials into "AI bills of materials" (AIBOMs) that record tool provenance (AHEAD).
• Reproducibility. None of the reviewed sources gives a benchmark, a detection rate or an open specification for lineage capture.

• "AI security is a data visibility problem." This is a sponsored argument with no data attached (MIT TR).
• Existing frameworks under-cover data leakage. This is supported independently: data leakage averaged 1.340 and non-determinism 1.231, the lowest category scores across the 16 frameworks (arXiv). The paper is a submission to a NIST request for information and has not been peer reviewed.
• 509% and 357% growth figures; 60% adoption of agentic apps on devices. These come from Cyberhaven Labs, and the press releases don't describe how they were measured (May release; Flow release). No independent source confirms them.

Data lineage is a long-standing data-governance practice, and data-catalog vendors already describe it as a basis for AI work (Atlan). Cyberhaven built its data-loss-prevention and insider-threat products on lineage before it pitched lineage for agents (VentureBeat). Identity-security commentators make a similar point: once agents can call tools or sub-agents, the key question becomes what an agent did, on whose authority, and in what order (NHI Mgmt Group).

• Cyberhaven sponsored the report and benefits from its conclusion. The full text is behind a download.
• The arXiv study finds that even the strongest framework, the OWASP Agentic Security Initiative, covers only 65.3% of the threats it lists (arXiv). Lineage addresses only part of that range.
• A competing vendor, Nightfall, publishes critical reviews of Cyberhaven (Nightfall). Those reviews are not neutral either.
• Capturing agents' prompts and file reads on employee devices creates a large store of sensitive data, which brings its own privacy and retention risks.

• Take an inventory of agents and MCP servers before setting policy. Shadow agents are the risk vendors keep pointing to.
• Make sure agent pipelines record provenance at each handoff. The report's pricing example fails because nobody can trace where restricted data entered the workflow.
• Check the compliance APIs of frontier-model vendors: Cyberhaven's integrations rely on Anthropic's and OpenAI's.
• Ask vendors for real detection and false-positive figures. None of the reviewed sources publishes them.