This piece covers a dozen announcements from a Solutions Review roundup, checked against the vendors' own releases. Nearly every capability claim comes from the vendor. None of it has been independently tested.
The headline theme: governing AI agents on endpoints
Tanium AgentSight: inventory now, enforcement later
Tanium announced AgentSight at its Converge conference in Las Vegas as a standalone product. Its own release says visibility is delivered today in the Tanium console, including on-premises and air-gapped environments. It finds AI components across Windows, macOS, and Linux, including local models, coding agents, MCP servers, desktop AI apps, IDE and browser extensions, and skills. Trade coverage adds that each discovered component is mapped to the device it runs on and the person who installed it. It also notes AI cost tracking through Tanium's Asset product, so IT and finance can see what AI software costs and where it is deployed.
The availability boundary is the part to read closely. Tanium's release describes fleet-wide control as work under development. That control would let teams set an AI policy once and remove or block unsanctioned AI in one action, with a detect-only mode first. Today you get discovery, inventory and cost tracking. The release lists some response actions, such as kill process, quarantine and Live Response, as available through Threat Response. Those are existing Threat Response features, not new AgentSight functions.
Tanium's release also carries the usual disclaimer that forward-looking functionality should not drive a purchasing decision. Take that seriously.
Airlock Digital 7.1: application control for agents
Airlock says Agentic Control shipped with v7.1 on November 17. It extends deny-by-default application control to actions that supported AI agents try to run. Per the company, each action is checked against customer policy before it executes. Prohibited actions are stopped, and the policy decision goes back to the agent. Administrators can add context or suggest an alternative so the agent can retry within policy.
- Supported agents at launch: Codex and Claude.
- Recording: sessions, commands, files, policy decisions, outcomes, risk activity, token usage and estimated cost.
- Review: teams can search sessions and read AI-generated summaries.
- Positioning: Airlock calls it a preventative check at the endpoint that complements AI platform controls, identity and access management, data security, and detection and response.
Airlock also cites an April 2026 Cloud Security Alliance report. It says 82% of organizations had unknown AI agents running and 65% had an agent-related security incident in the prior 12 months. Treat those as survey figures Airlock chose to quote, not universal measurements.
Cato Networks AgentSense: stopping injected instructions
Cato's AgentSense is a runtime capability meant to stop harmful agent actions before they execute. Its Untrusted Influence Detection feature separates trusted instructions from untrusted content and traces how that content may steer an agent. The target is indirect prompt injection, including attacks that don't match known patterns. Cato's page wasn't retrievable for this report, so these details come from the Solutions Review summary. I found no general-availability date or independent efficacy data.
DTEX Agentic Defenders: mostly previews
DTEX's update shows how easily "launch" language blurs. Precision Tuner is in private preview. It evaluates security performance and recommends changes to detection logic, enrichment profiles, scoring models and policies, with analysts approving the changes. Triage Guardian, Threat Hunter and Deep Researcher are in Early Access.
DTEX says its expanded AI-interaction visibility captures prompts and responses without TLS inspection or browser extensions. It links them to endpoint activity, process lineage, file access and data movement. New playbooks include hunts for "lethal trifecta" agents, meaning private-data access plus untrusted content plus external action ability. Another covers agents coordinating through improvised channels. For availability of the playbooks and the visibility feature, DTEX tells customers to contact the company.
Exaforce ExaU and Expel Workbench: AI in the SOC
Exaforce launched ExaU, which lets individual practitioners create persistent AI agents from templates or plain-language instructions. CEO Ankur Singla describes the model as building your own agents and telling them what you're responsible for. The company claims the agents keep priorities and decisions and can be shared with colleagues. No performance measurements accompany that claim.
Expel's Workbench redesign is a staged rollout, per its release:
- Live today: the Precedent Engine, rapid triage agents, a deep investigation agent and a detection engineering agent.
- Early access: Cases, which show disposition logic and a reasoning path for each call.
- Late 2026: Entra ID sync, to bring users and groups into the context engine.
- Q1 2027: Workflow Builder, generally available.
- Early 2027: an MCP server, in early access.
Entra ID sync is the item most relevant to Microsoft-centric shops. Expel says it will extend context to customers' directories and sync users and groups automatically.
Metano: new entrant, generally available
Metano emerged from stealth on November 19 with $12 million in seed funding led by Sierra Ventures. Correlation Ventures, High Alpha and 360 ONE also took part. The company says the platform discovers agents, MCP servers, skills and packages. It uses a context graph called Hive to judge whether an agent's session behavior matches its purpose, and it enforces guardrails where agents run. Metano says the product is generally available and offers a free 30-day Eval Edition. It also says its detections map to MITRE ATLAS and the OWASP Top 10 for Agentic Applications. Those are company claims.
Exposure management and OT compliance
Cognizant joins the Qualys mROC Alliance
Cognizant will pair its vulnerability management services with Qualys Enterprise TruRisk Management and TruRisk Eliminate. The managed model has three stages:
- Detection: Qualys InstaScan identifies affected assets from existing inventory data.
- Prioritization: Agent Val validates which vulnerabilities are exploitable in each environment, and risk quantification estimates financial impact.
- Remediation: TruRisk Eliminate picks a patch, mitigation or isolation, verifies it, and rolls back if system health declines.
Cognizant works within each client's existing tools, change windows and approval processes. It phases in automation starting with low-risk systems.
The release makes striking claims. It says Microsoft's November Patch Tuesday disclosed 1,093 vulnerabilities in one release, up from 93 before frontier AI models arrived. It says InstaScan delivered 90% of detections within minutes. It also says only 357 of 48,172 vulnerabilities disclosed in 2025 (0.74%) were confirmed remotely exploitable and weaponized. These are Qualys figures from a commercial announcement, not independently verified. Windows admins should check the Patch Tuesday count against Microsoft's own release notes before repeating it. The exploitability point is still a sensible one: patch order should follow real exposure, not raw CVE counts.
Nozomi Networks and Lockheed Martin: STIG checks for OT
The two companies announced a co-development project, funded as an offset tied to the F-35 sale to Switzerland. The first deliverable is automated security configuration assessment mapped to DISA STIGs. Nozomi says that existing STIG workflows suit laptops and servers but fit engineering workstations and facility systems poorly. Many defense organizations still lean on spreadsheets and manual validation.
Nozomi says Arc sensors collect host-level configuration data. Results are managed in the Central Management Console or in Vantage or Vantage for Government. Planned future work includes active polling for assets that can't run sensors and support for more frameworks. The release frames these as capabilities being built, so don't read them as shipping features.
Network architecture and supply chain
Orange Business Alternative WAN
Orange plans customer proof-of-concept projects for Alternative WAN, a cloud-centric design positioned alongside SD-WAN. It moves network intelligence into Orange's Evolution Platform so branches can run standardized hardware. It was developed with the prpl Foundation. Per the plan, proofs of concept start in mid-2027, with a second development phase in late 2027. Cost and emissions reductions, including target cuts of 30% in Scope 3 and 90% in Scope 2 emissions, are projections against a reference architecture, not customer results.
StepSecurity: funding plus a free registry tier
StepSecurity raised a $20 million Series A led by Elephant, for $23 million in total funding. It also introduced a free Secure Registry community tier integrated with Harden-Runner. That lets public repositories route package retrieval through the registry and block compromised packages. The company says more than 20,000 open-source projects use Harden-Runner's free tier. This is the most directly usable item for developers on GitHub Actions, since the free tier needs no procurement cycle. Check the scope before relying on it, because the release describes it for public repositories.
Strike Graph Integration AI
Strike Graph's Integration AI builds API-based connections that collect compliance evidence continuously, beyond a fixed catalog. The company says integrations are read-only by default and carry endpoint allowlists validated against vendor API documentation. It also says credentials are kept in a managed secrets vault, not handled during construction. These are design claims from the vendor.
What to take from the week
| Item | Status per vendor |
|---|---|
| Airlock Agentic Control | Available in v7.1 (Codex, Claude) |
| Tanium AgentSight | Visibility available; fleet-wide control in development |
| Metano | Generally available; free 30-day eval |
| DTEX Precision Tuner | Private preview |
| Expel Workbench | Staged through 2027 |
| Nozomi/Lockheed STIG automation | Co-development project |
| Orange Alternative WAN | Proofs of concept from mid-2027 |
A few practical observations, based on general industry knowledge rather than vendor claims:
- Inventory comes before policy. Whichever tool you pick, you can't govern AI software you haven't found. Discovery of shadow AI is the part that works today.
- Ask what "block" means. Airlock acts on individual agent actions. Tanium's planned control acts on installed software. Cato and Metano watch agent behavior. These are different layers and can overlap.
- Agent support is narrow. Airlock names Codex and Claude. Check any vendor's list against the agents your developers use.
- Preview is not production. Several of this week's items have no firm ship date, so keep them out of near-term budgets.
- Vendor statistics are marketing inputs. Survey percentages, detection-speed claims and emissions targets all came from the sellers.
If you run a Windows fleet, the useful move is a pilot of discovery. Find out which local models, MCP servers and IDE extensions are already installed before deciding what to forbid.
References
- Endpoint Security and Network Monitoring News for the Week of November 20th: Tanium, Exaforce, Metano, and More - Solutions Review Solutions Review · 2026-11-20T14:44:07+00:00
- Metano Emerges from Stealth to Unify AI Agent Security metano.ai
- DTEX Introduces Precision Tuner for Continuous AI Security Optimization - DTEX dtex.ai