Palo Alto Bets Agentic Endpoint Security Is the Next Endpoint Category
Cortex XDR One merges XDR with the technology Palo Alto bought from Koi, and treats an AI coding agent with the same suspicion as a user holding domain admin.
Palo Alto Networks released Cortex XDR One, which merges its endpoint detection and response platform with the agentic endpoint security it acquired from Koi in April, and the pitch is unusually blunt. An AI coding agent running on a laptop with a developer’s credentials is, in the company’s own framing, the ultimate insider threat. It reads files, runs shell commands, calls tools and moves data, and it does so non-deterministically under permissions nobody granted deliberately.
The product is available now, both as the combined XDR One platform and as a standalone module that runs beside a third-party EDR through a targeted sensor. Alongside it, Palo Alto shipped Cortex XDR 6.0, which adds operating-system-level interception of prompts and responses, preconfigured policies mapped to the MITRE ATLAS framework, small language models running on the endpoint sensor, and AI-assisted YARA rule management.
What actually changed
Two design decisions carry most of the weight. The first is where the interception happens. Many AI security tools work through hooks in the applications they support, which means they see Claude Code or Cursor but not the next agent to appear. Palo Alto now intercepts prompts and responses at the operating system level, which extends coverage beyond apps with native monitoring hooks. It names Claude Code, Codex CLI and open-source frameworks such as OpenClaw as covered, and the architecture implies the list can grow without waiting for each vendor to expose an interface.
The second is where the analysis runs. Optimised small language models sit on the endpoint sensor and inspect scripts before execution, which matters for two reasons. It works offline, and it keeps source code and file contents on the machine rather than shipping them to a cloud service for inspection. Palo Alto claims no additional deployment overhead or performance degradation, which is the kind of claim a security team will test on its own fleet before believing.
Two smaller additions matter more than they look. The preconfigured policies map to MITRE ATLAS, the framework that adapts ATT&CK techniques to adversarial machine learning, which gives an administrator a starting point instead of a blank policy editor, and gives a security team a shared vocabulary when it reports to a board. The AI-assisted YARA management goes further: Palo Alto’s AgentiX can draft detection rules, estimate whether they will generate excessive alerts, and search the endpoint estate for matching files without an analyst connecting to machines one at a time. Rules are the cheapest currency in detection engineering, and making them easier to write is a modest but real productivity gain.
Why agents break the old model
The argument for agentic endpoint security rests on a specific technical observation: modern endpoint behaviour no longer comes only from executables. Code packages, browser extensions, IDE plugins, scripts, local MCP servers, containers and model artefacts all run on a machine, employees install most of them without central oversight, and traditional endpoint tooling was built to watch binaries. An agent compounds this because it acts with a real user’s credentials, can install further components at machine speed, and behaves probabilistically. Prompt injection, indirect data poisoning and unauthorised API calls follow from that combination, and neither a container nor a prompt-level filter closes the gap.
The shadow AI problem has a shape the release only gestures at. Employees sign into coding agents with personal accounts on corporate machines, and that activity sits outside organisational control. Palo Alto detects it by comparing the account’s email domain with the company’s, and flags such sessions across Claude Code, Cursor, Codex and Antigravity. The risk is not only the files an agent reads. Free and personal subscription tiers often use conversation data for model training by default, while enterprise agreements carry no-training commitments, so a developer on a personal account can move proprietary code into a training pipeline without anyone deciding to. That is a governance failure no firewall catches, and it is the number a chief information security officer will want before approving wider agent rollout.
The Koi origin
The most interesting part of the story is how the technology arrived. Palo Alto’s own information security team found Koi while the startup was still in stealth mode, deployed it to protect the company’s internal AI workloads, and the company then completed the acquisition in April 2026. That is a cleaner thesis than most acquisitions carry: the buyer was the first customer, and it tested the product on its own engineers before selling it. Koi’s chief executive, Amit Assaraf, and Palo Alto’s chief product and technology officer, Lee Klarich, both described agents as the ultimate insider threat at the time of the deal, and the language has not changed since. The deal also explains the product’s shape. Koi shipped discovery, runtime policy and remediation features through 2026, and XDR One consolidates that work into a single console rather than starting again.
The threat record it points at
Agentic endpoint security is not a hypothetical category, and Palo Alto’s own material lists the incidents it means. GlassWorm, in December 2025, put 50,000 malicious downloads into a public marketplace. Shai-Hulud 2.0 compromised around 1,000 new code repositories every 30 minutes in November 2025. ChainDrop in August 2026 infected 400 popular developer packages and harvested credentials. Each of those attacks targets the supply chain an agent consumes rather than the agent itself, which is why the product pairs runtime guardrails with inventory. You cannot block what you have not catalogued.
Remediation is the other half, and the harder one. Discovery tells a team what is running; removal is what closes the risk. Palo Alto describes deleting a flagged plugin’s entire footprint from a machine, including user settings, the local plugin directory and associated files, and clearing npm packages installed during a campaign such as Shai-Hulud before blocking re-introduction through a registry proxy. Doing that without breaking a developer’s environment is the real test, and the feature set suggests the company learned it from its own internal deployment.
The claims that need testing
Three claims deserve scrutiny before a buyer signs. The first is the false-positive reduction of up to 95 per cent, which Palo Alto attributes to stitching cross-domain telemetry into end-to-end attack chains. That is a vendor figure with no published methodology, and false positives are precisely what determines whether analysts trust an alert queue. The second is that on-sensor inference carries no performance penalty. Every security agent on a developer laptop competes for CPU with the work the laptop exists to do, and the honest answer is usually measurable but acceptable. The third is coverage. Operating-system-level interception is a stronger claim than application hooks, but prompt injection is an adversarial problem, and no vendor in this category has published detection rates against a standard corpus.
The enforcement question
The design choice with the most practical consequence is how the product handles a blocked action. Palo Alto offers two modes: block, which denies outright, and ask, which pauses the agent and requests explicit user approval. The second is the interesting one, because the failure mode of agent security is not a missed block but a policy that makes engineers stop using the tools. A system that interrupts a developer mid-task to approve a shell command will be tuned down or bypassed within a week. Palo Alto’s answer, a block message inside the agent’s own chat interface with an in-line route to request an exception, reads as an attempt to keep the friction where the developer already is rather than in a separate console.
The competitive and consolidation picture
Palo Alto is not alone in agentic endpoint security. CrowdStrike, Microsoft and SentinelOne have all extended endpoint platforms toward AI workloads, and cloud security vendors argue that agent risk belongs in the identity and data layer rather than on the host. Palo Alto’s differentiator is consolidation: it wants agent telemetry on the same sensor as everything else, feeding Cortex XSIAM, so that one platform holds the picture. The standalone option exists because most large enterprises already run someone else’s EDR, and asking them to remove it in order to buy an agent module would stall the sale before it started.

What it signals
The release marks a category graduating from concept to product line. Eighteen months ago, securing AI agents on endpoints was a research topic and a handful of startups. Now a major vendor sells it as a module with a console, a managed service behind it, and policies mapped to a public threat framework. That is progress, and it also moves the hard questions from whether this can be built to whether it works. The answers will come from customers willing to publish detection rates, and from incidents that happen anyway. For a security leader, the practical advice is unglamorous: inventory first, because the shadow AI number will be larger than expected, then enforce in ask mode, then measure what analysts actually action before switching anything to block.
One further point is worth keeping in view. How much autonomy an enterprise grants an agent is a management decision, and no sensor makes it for you. The other thing to watch is whether the category survives contact with procurement: a module that must run beside someone else’s EDR competes for budget with that EDR’s own AI features, and the standalone option is an admission that many buyers will start there.

