# Agentic AI in Cybersecurity > What AI agents do in security work, job by job: what they read, what they do without a person, where a person approves and what evidence they leave. Edition 2026-11-19. Data under CC BY 4.0: https://agenticaiincybersecurity.com/data/use-case-matrix.csv, https://agenticaiincybersecurity.com/data/evaluation-checklist.csv. Full text: https://agenticaiincybersecurity.com/llms-full.txt ## Use cases and vendors - [Agentic AI in Cybersecurity: 85 Named Agents, 39 Vendors](https://agenticaiincybersecurity.com/index.md): Pick one of six security jobs and see what agents read, do alone, who approves and what they leave behind. 85 named agents, 39 vendors, read 19 Nov 2026. - [AI Agents for Access Reviews: Approvals and Evidence](https://agenticaiincybersecurity.com/use-cases/access-reviews/index.md): What Lumos and Veza say their access review agents do, who signs the certification, the record to keep, and a worked campaign from 1 to 30 Oct 2026. - [AI Agents for Cloud Posture: Approvals and Evidence](https://agenticaiincybersecurity.com/use-cases/cloud-posture/index.md): Seven vendors publish a cloud posture agent, from read-only to automatic fixes. Approval points, the record to keep, and a worked bucket finding of 3 Nov 2026. - [AI Agents for Code Fixes: Who Merges, What to Keep](https://agenticaiincybersecurity.com/use-cases/code-and-dependency-fixes/index.md): 16 vendors publish an agent that proposes code or dependency fixes. Who approves, the pull request record, and a worked example on CVE-2026-94545 (22 Sep 2026). - [AI Agents for New-Disclosure Response: Approvals and Records](https://agenticaiincybersecurity.com/use-cases/new-disclosure-response/index.md): 10 vendors publish an agent that matches a new CVE to your assets. Approval points, the record to keep, and CVE-2026-111497 followed from 14 Oct to 10 Nov 2026. - [AI Threat Modeling Agents: 6 Vendors and What to Keep](https://agenticaiincybersecurity.com/use-cases/threat-modeling/index.md): 6 vendors publish an agent that drafts a threat model from prompts, diagrams or a repository. Acceptance record to keep, and a worked example on CVE-2026-9198. - [AI Agents for Vulnerability Triage and Remediation](https://agenticaiincybersecurity.com/use-cases/vulnerability-triage-and-remediation/index.md): 28 vendors publish an agent that triages scanner findings or moves them toward a fix. Approval points, the record kept, and a worked example on CVE-2026-82329. ## Agent Evaluation Checklist - [Agent Evaluation Checklist: 20 Questions and Tests](https://agenticaiincybersecurity.com/evaluation-checklist/index.md): Twenty questions in 7 areas to settle before an AI agent changes production, each with a test you run on your own systems and the guidance line behind it. - [AI Agent Audit Trail: A Log Entry You Can Reconstruct](https://agenticaiincybersecurity.com/evaluation-checklist/evidence-and-audit-trail/index.md): Checklist question Q09 with a complete example log entry for one agent change, five reconstruction questions and a retention period worked to a date. - [Human Approval Points: Checklist Questions Q05, Q06, Q08](https://agenticaiincybersecurity.com/evaluation-checklist/human-approval-points/index.md): Where an agent must pause for a person: the approval map to ask for, the risk rating that drives it, and a test that shows whether approvers decide or click. - [Kill Switch and Rollback: Checklist Questions Q12 and Q13](https://agenticaiincybersecurity.com/evaluation-checklist/kill-switch-and-rollback/index.md): What to ask for before an agent changes production: who can stop it, how each change is undone, and an incident plan with switch-off criteria and a drill. - [Does the Security Job Need an Agent? Checklist Question Q01](https://agenticaiincybersecurity.com/evaluation-checklist/need-an-agent-at-all/index.md): Q01: split one real finding, CVE-2021-44228, into fixed and variable steps and check whether a workflow with one model call does the job before an agent. - [Re-Evaluate After Model Changes: Questions Q16 and Q11](https://agenticaiincybersecurity.com/evaluation-checklist/re-evaluate-after-model-changes/index.md): Which changes re-run the test set, what result blocks release, and how to catch a supplier-side change you never scheduled, with a worked run of 48 cases. - [Scope of Change: What an AI Agent May Touch, Per Tool](https://agenticaiincybersecurity.com/evaluation-checklist/scope-of-change/index.md): Checklist Q02 and Q03: give each tool one system and one of six change kinds (read to delete), then replace a shell tool using 412 logged calls over 30 days. - [Stop and Handover Conditions for a Security Agent](https://agenticaiincybersecurity.com/evaluation-checklist/stop-and-handover-conditions/index.md): Checklist Q07: the retry, iteration and rate limits that halt an agent, a filled-in handover note, and a stop on CVE-2026-123158 timed from 09:21 to 09:38. - [Success Criteria and Test Sets for a Security Agent](https://agenticaiincybersecurity.com/evaluation-checklist/success-criteria-and-test-sets/index.md): Checklist Q14 and Q15: a criteria template, a 40-case test set run 3 times (109 of 120 trials passed) and why that still failed, with sandbox checks. - [Supplier Controls: Questions to Ask AI Agent Vendors](https://agenticaiincybersecurity.com/evaluation-checklist/supplier-controls-when-buying/index.md): Checklist Q19 and Q20: nine questions for an AI agent supplier, each with an acceptance test and contract wording, and a one-day re-test window worked out. - [Tool Permissions and Identity: What an Agent Can Do](https://agenticaiincybersecurity.com/evaluation-checklist/tool-permissions-and-identity/index.md): Checklist Q04: which identity and permissions a security agent uses on each system, with a worked AWS read-only policy, a refused-action log line, review dates. - [Untrusted Input: What Outside Text Can Make an Agent Do](https://agenticaiincybersecurity.com/evaluation-checklist/untrusted-input/index.md): Checklist Q18: which text a security agent reads comes from outside your control and which tools stay active then, with a harness rule and an 18-alert replay. - [Verify After Each Change: Check the System, Not the Agent](https://agenticaiincybersecurity.com/evaluation-checklist/verify-after-each-change/index.md): Checklist Q10: the separate check that confirms an agent's change happened, worked through on a Next.js upgrade for CVE-2026-75604, and how each check can fail. ## Agent harness and benchmarks - [What Is an Agent Harness? Eight Definitions, One CVE](https://agenticaiincybersecurity.com/agent-harness/index.md): Eight published agent harness definitions, quoted and dated, and a worked case: CVE-2026-82329 triaged by one model in two harnesses, with two results. - [Agent Harness Approvals: Six Mechanisms, One Traced Call](https://agenticaiincybersecurity.com/agent-harness/approvals-and-checkpoints/index.md): How six agent harnesses pause a tool call for approval (versions read 19 Nov 2026), the OpenAI shell defaults, and a CVE-2026-0322 change traced to its log. - [Agent Harness Memory: Six Designs and a Checklist File](https://agenticaiincybersecurity.com/agent-harness/context-and-memory/index.md): How six agent harnesses keep memory (versions read 19 Nov 2026), plus a CVE-2026-9198 checklist file with agent and human fields and a retention table. - [Agent Eval Harness: A Worked Security Example](https://agenticaiincybersecurity.com/agent-harness/evaluation-harness/index.md): Eval harness for a triage agent: five tasks from CVE-2026-82329, CVE-2026-41709 and CVE-2026-9198, with graders, three trials each, and pass@3 against pass^3. - [How to Build an Agent Harness: Steps, Owners, Records](https://agenticaiincybersecurity.com/agent-harness/how-to-build-an-agent-harness/index.md): Eight steps to build a security agent harness, each with an owner, the record it leaves and how it fails, worked on a triage harness for CVE-2026-82329. - [How to Test an Agent Harness: A Worked Test Run](https://agenticaiincybersecurity.com/agent-harness/how-to-test-an-agent-harness/index.md): Eight harness tests tied to documented gaps in Claude Agent SDK 0.2.185, OpenAI Agents SDK 0.25.0 and Deep Agents 0.7.40, run on an Artifactory response agent. - [Agent Harness Logging: A Filled Evidence Record](https://agenticaiincybersecurity.com/agent-harness/logging-and-evidence/index.md): A filled evidence record for one agent action on Artifactory CVE-2026-82329, with the logging each SDK gives by default and why the log must live off the host. - [One Agent or Many: Orchestration for Security Agents](https://agenticaiincybersecurity.com/agent-harness/orchestration-and-subagents/index.md): When a security agent needs subagents: Slack's Director, Critic and four experts, a worked token bill at Anthropic's 4x and 15x figures, and SDK limits. - [Tool Permissions for a Security Agent Harness](https://agenticaiincybersecurity.com/agent-harness/tools-and-permissions/index.md): How six agent harnesses decide whether a tool call runs, five documented gaps to test, and a filled register for four security tools. Docs read 19 Nov 2026. - [Security AI Agent Benchmarks: Results and Method Table](https://agenticaiincybersecurity.com/benchmarks/index.md): Six benchmark families in one table: what each measures, how it grades, and its latest published results with model versions and dates, read 19 Nov 2026. - [BountyBench Results: Detect, Exploit and Patch by Agent](https://agenticaiincybersecurity.com/benchmarks/bountybench/index.md): BountyBench results for 10 agents on 120 tasks from 40 bug bounties, with model versions, attempts and bounty dollars, and how its invariant check works. - [CTIBench Results: Threat Intel Tasks for LLMs](https://agenticaiincybersecurity.com/benchmarks/ctibench/index.md): CTIBench (arXiv 2406.07599): 2,500 questions and four CTI tasks, with results for five models, sample sizes worked into intervals, and how teams use it. - [CVE-Bench Results: Agents Exploiting 40 Web CVEs](https://agenticaiincybersecurity.com/benchmarks/cve-bench/index.md): CVE-Bench (arXiv 2503.17332): 40 critical web CVEs, eight graded outcomes, published success rates, token costs and failure modes for three 2025 agents. - [CyberGym Results: Reproducing 1,507 Real Bugs](https://agenticaiincybersecurity.com/benchmarks/cybergym/index.md): CyberGym (arXiv 2506.02548): agents reproduce 1,507 OSS-Fuzz bugs from a description. Four levels, the paper's baselines and the project table to Oct 2026. - [SWE-bench Results: Dated Percentages and What They Mean](https://agenticaiincybersecurity.com/benchmarks/swe-bench/index.md): SWE-bench results from swebench.com on 19 Nov 2026: 13 models on Verified and Multilingual, Full history since 2023, and the Multimodal 480 vs 517 count. ## Guides - [Agent Harness vs Agent Framework: Which Layer to Pick](https://agenticaiincybersecurity.com/guides/agent-harness-vs-agent-framework/index.md): Runtime, framework or harness for a security agent? One job, the Cisco ISE patch check, worked through four routes, with what the team must still configure. - [How Long to Keep AI Agent Run Logs and Which Fields](https://agenticaiincybersecurity.com/guides/agent-run-logs-what-to-keep/index.md): EU AI Act Articles 19 and 26 set six months for high-risk systems from 2 Dec 2027. A keep, store or drop table per agent log field, plus a dated change record. - [Agentic AI vs Generative AI in Security Work Compared](https://agenticaiincybersecurity.com/guides/agentic-ai-vs-generative-ai-in-security/index.md): The same Atlassian advisory, CVE-2026-21589, handled as a one-shot summary and as an agent with tools: what changes in output, error cost and testing. - [Agentic AI vs SOAR Playbooks: When a Playbook Is Enough](https://agenticaiincybersecurity.com/guides/agentic-ai-vs-soar-playbooks/index.md): A SOAR playbook fixes its branches in advance; an agent picks its next step. On CVE-2026-0322: 20 fixed builds, one step for an agent, four for a playbook. - [AI Agent Evals: Grading Runs and Setting a Release Gate](https://agenticaiincybersecurity.com/guides/ai-agent-evals/index.md): Grade an agent eval with pass@1, pass@3 and pass^3 on 40 findings and 120 trials, then compare the result with a gate written before the run (19 Nov 2026). - [AI Agent Guardrails: A Graded Tool List With Limits](https://agenticaiincybersecurity.com/guides/ai-agent-guardrails/index.md): Build guardrails for a security agent as a graded tool list: nine tools for CVE-2026-111497 rated low, medium or high, each with access, undo and approval rule. - [AI Agent Autonomy Levels: Rating Eight Security Actions](https://agenticaiincybersecurity.com/guides/autonomy-levels/index.md): Set autonomy per action, not per product: eight security actions rated on the NCSC's five dimensions, mapped to the five levels of Feng, McDonald and Zhang. - [Build vs Buy AI Agents for Security: What Decides It](https://agenticaiincybersecurity.com/guides/build-vs-buy-ai-agents/index.md): Decide build or buy for a security agent on four tests before cost: a product for the job, an approval gate you set, exportable logs, a swappable model. - [Choosing a Model for a Security Agent: Limits and Dates](https://agenticaiincybersecurity.com/guides/choosing-a-model-for-a-security-agent/index.md): Context limits, retirement dates, retention terms and a worked 120-trial comparison of two models, read from Anthropic, Google and OpenAI pages on 19 Nov 2026. - [Cost of Running Security Agents: Worked Per-Task Figures](https://agenticaiincybersecurity.com/guides/cost-of-running-security-agents/index.md): Token cost of two security tasks at twelve published price points, then caching, batch, retries and review time worked through. Prices read 19 Nov 2026. - [Evaluate a Security Agent on Your Own Findings](https://agenticaiincybersecurity.com/guides/evaluate-a-security-agent-on-your-findings/index.md): A worked trial of a dependency-fix agent on 40 findings: four checks in a clean copy, a 52 to 80% interval on 27 passes, thresholds written first, the record. - [How to Use Agentic AI in Cybersecurity: Six Steps](https://agenticaiincybersecurity.com/guides/how-to-use-agentic-ai-in-cybersecurity/index.md): Six steps to put an agent on one security job, with the owner, the record and the usual failure for each, NCSC risk levels and a dated rollout worked through. - [Designing the Approval Step for a Security Agent](https://agenticaiincybersecurity.com/guides/human-approval-design/index.md): What an approver needs to judge an agent's change: a filled approval record for CVE-2026-75604, how many accepted items to sample, and expiry rules. - [Human in the Loop vs Human on the Loop in Security](https://agenticaiincybersecurity.com/guides/human-in-the-loop-vs-human-on-the-loop/index.md): In the loop means a person approves before the action; on the loop means a person supervises after. Three actions from CVE-2026-123158 placed in each mode. - [Measuring a Security Agent After It Ships](https://agenticaiincybersecurity.com/guides/measuring-an-agent-after-it-ships/index.md): Six production metrics for a security agent with formulas, thresholds, the action each triggers, a worked four-week review and Rival's block-rate series. - [When a Security Task Needs More Than One Agent](https://agenticaiincybersecurity.com/guides/multi-agent-systems-in-security-work/index.md): Split a security agent only for tool overload or privilege. OpenAI's 15 and 10 tool counts, LangChain's call and token counts, and a worked split of 14 tools. - [What a Security Agent Should Never Do Alone](https://agenticaiincybersecurity.com/guides/what-an-agent-should-never-do-alone/index.md): Seven actions a security agent must not take without a named person, graded on the NCSC 0 to 4 scale with Red Hat's CVE-2026-46300 fixes and a real failure. - [What Is Agentic AI in Cybersecurity? A Working Definition](https://agenticaiincybersecurity.com/guides/what-is-agentic-ai-in-cybersecurity/index.md): Agentic AI in cybersecurity defined from Anthropic, OWASP and LangChain, then run on CVE-2026-46300 as a workflow and as an agent, with tests for a false label. - [Writing a Task Spec for a Security Agent: Template](https://agenticaiincybersecurity.com/guides/writing-a-task-spec-for-a-security-agent/index.md): A complete eight-field task spec for a security agent, filled in for CVE-2026-46300 with real kernel versions, a 31 Oct 2026 due date and sourced stop rules. ## Reference - [Agentic AI in Cybersecurity News: 14 Dates, Sep to Nov 2026](https://agenticaiincybersecurity.com/dates-to-know/index.md): Fourteen dated items from 21 Sep to 18 Nov 2026 that matter to a team running security agents, each read from its publisher on 19 Nov 2026 and labeled. - [Agentic AI Security Terms: 40 Defined, 19 Quoted](https://agenticaiincybersecurity.com/glossary/index.md): 40 terms for teams running AI agents on security work: 19 quoted from named issuers such as NIST CSRC, OWASP and MITRE ATLAS, 21 in the site's own words. ## About the site - [About Agentic AI in Cybersecurity: Scope and Hosting](https://agenticaiincybersecurity.com/about/index.md): What the site covers (agents doing security work), what it leaves out, how it is hosted on Amazon S3 and CloudFront, and its CC BY 4.0 license. 19 Nov 2026. - [Contact: Corrections, Missing Vendors and Questions](https://agenticaiincybersecurity.com/contact/index.md): One address for corrections, vendor pages we missed and data requests, what to include so a message can be acted on, and where vulnerability reports go. - [Corrections Log: Old and New Values, Dated](https://agenticaiincybersecurity.com/corrections/index.md): Every change to a fact, quotation, number or date on this site, with the old value, the new value and the date, plus the pages merged on 19 Nov 2026. - [Privacy: What This Site Stores and What It Does Not](https://agenticaiincybersecurity.com/privacy/index.md): The site sets no cookies and loads no third-party scripts. Access logs are deleted after 400 days and emails after 12 months. Data requests: 30 days. - [Terms of Use: Reuse, Quotations and Limits](https://agenticaiincybersecurity.com/terms/index.md): Terms for using the site: text and data under CC BY 4.0 with a cite line, quotations stay with their authors, no warranty, and the security content rule.