EDR sees the file write, not the reason: the case for an agent-native detection layer

A team of Uber-affiliated authors and academic collaborators has published the design and production record of ADR, a detection and response system for AI agents operating through the Model Context Protocol, carrying a copyright line for the MLSys Conference Industry Track in Bellevue, Washington, 2026 ^[1]. Their central claim is structural rather than incremental: endpoint detection and response tools can observe a file write or a network call, but not the user prompt, the agent's reasoning steps, or the causal chain that links intent to that execution ^[2].

That is a real gap, and it is not one you close with tuning. MCP is a standardised interface for connecting large language models to external tools and data sources ^[3], and the paper argues it opens an attack surface where agents can be manipulated through natural language, exploited via compromised MCP servers, or coerced into running unsafe commands and exfiltrating data ^[4]. An EDR alert on a suspicious write tells you a process misbehaved; it does not tell you whether a poisoned tool description talked the agent into it. The authors also fault static guardrails and rule-based policy checkers for failing to generalise across attack techniques and enterprise contexts ^[5], and flag severe class imbalance as a further obstacle ^[6].

The third constraint is the one operators will feel first: the paper states that LLM-based inference is prohibitively expensive at scale ^[7]. ADR's answer is a three-part split. A lightweight endpoint sensor reconstructs high-fidelity agentic telemetry ^[8]; an "Explorer" component does pre-deployment red teaming and hard-example generation ^[9]; and the detector runs two tiers, fast triage in front of context-aware reasoning ^[10]. In other words, you do not get to put a reasoning model on every session, so you build a cheap filter and reserve the expensive judgement for what survives it.

The deployment numbers are the most interesting part, because they are operational rather than benchmark. ADR has run for more than ten months across over 7,200 unique hosts, processing more than 10,000 agent sessions daily ^[11]. It surfaced hundreds of credential exposures across 26 categories, and the authors report a shift-left prevention layer at 97.2 percent precision with 206 detected credentials ^[12]. Credential leakage, not exotic agent hijacking, is what the telemetry actually found at volume.

The benchmark results are more mixed than the framing suggests. On ADR-Bench, which the same team introduces with 302 tasks, 17 techniques and 133 MCP servers, ADR reports zero false positives while detecting 67 percent of attacks, which the paper says beats ALRPHFS, GuardAgent and LlamaFirewall "by 2-4 in F1-score" ^[13]. Zero false positives is worth something; missing roughly a third of attacks on your own benchmark is worth remembering ^[14]. On the public AgentDojo prompt injection benchmark it detects all attacks with three false alarms across 93 tasks ^[15], about 3.2 percent of tasks ^[16].

Two caveats. These are self-reported figures from a single enterprise deployment, and the ADR-Bench scoreboard was built by the party being scored ^[13]. The version available is an arXiv rendering carrying automated warnings that its page layout violates the ICML style ^[17].

Worth watching: whether anyone outside Uber reproduces the 67 percent figure, whether the 97.2 percent precision on credential prevention holds as agent traffic grows past 10,000 daily sessions, and whether ADR-Bench is actually released in a form competitors can run.