
Security Operations & Response
AI Cyber Defense
ColdAI's Cyber Defense practice uses frontier AI models, through access programs for vetted defenders, to find and fix vulnerabilities, triage security alerts and prepare incident response for organizations defending systems they own. Engagements start with a scoped exposure assessment and run as a fixed sprint or a managed service, and a person approves every change to a production system.
What does ColdAI do in AI Cyber Defense?
AI cyber defense services use frontier language models alongside conventional scanners and human analysts to find exploitable weaknesses, triage security alerts, prepare for incidents and secure the AI systems an organization deploys. ColdAI works only on systems you own or are authorized to test, starts with a scoped exposure assessment, and has a person approve every change to a production system.
Frontier model access
Frontier AI, cleared for defense
OpenAI and Anthropic each run a program that gives vetted security teams access to frontier models for defensive work.
OpenAI
Daybreak
ColdAI is part of OpenAI's Daybreak program for cyber defenders.
OpenAI's program for cyber defense, opened in stages to vetted security organizations.
OpenAI: Daybreak (opens in a new tab)Anthropic
Cyber Verification Program
ColdAI is verified under Anthropic's Cyber Verification Program, with access to Claude Mythos for defensive security work.
Anthropic's verification program for cyber defenders, which gives vetted security teams access to its frontier models, including Claude Mythos, for defensive work.
Anthropic: Cyber Verification Program (opens in a new tab)
Access is granted by each provider for defensive security work. It is not an endorsement of ColdAI's services by OpenAI or Anthropic.
What does ColdAI's AI Cyber Defense service include?
- AI-assisted vulnerability discovery
- SOC alert triage and enrichment
- Incident response readiness
- Security for AI systems and agents
In depth
Method pages, checklists and a buyer's guide
Read how each workstream works before you decide what to scope.
- 01ProcessAI Vulnerability Discovery: How an Engagement RunsHow an AI-assisted vulnerability discovery engagement runs: rules of engagement, model-assisted review, human reproduction, fix and retest, and reporting.
- 02ArchitectureAI SOC Alert Triage: Reference ArchitectureA reference architecture for AI SOC alert triage and enrichment: data sources, model placement, guardrails, analyst sign-off, evaluation and drift monitoring.
- 03Deep diveFrontier AI Models for Cyber Defense: Access ProgramsFrontier AI models for cyber defense: how Anthropic's Cyber Verification Program and OpenAI's Daybreak verify defenders, their tiers and conditions.
- 04ChecklistExternal Attack Surface Assessment ChecklistA working checklist for an external attack surface assessment: authorization, asset discovery, weakness checks, prioritization with KEV and EPSS, and retest.
- 05Buyer's guideManaged Detection and Response Buyer's GuideA neutral managed detection and response buyers guide: what to compare, questions to ask any provider, how AI is used, and red flags before you sign.
How an engagement runs, from first scope to a verified fix
A short sprint and an ongoing managed service share the same control points. What differs is how long the work stays open.
- Scope and authorize
The client lists the systems in scope and signs rules of engagement. Testing touches only systems the client owns or is authorized to test.
- Exposure assessment
A scoped first piece of work that maps what an attacker can reach and ranks it, so the next step is chosen from evidence.
- Sprint or managed service
A fixed-scope sprint for a defined question, or a managed service when the work repeats every week.
- Human reproduction
A security engineer reproduces every model-assisted finding before anyone is asked to act on it.
- Named approver
A person at the client approves each change to a production system, and the approval is recorded.
- Retest and report
Each fix is retested. The report records what was found, what changed and what remains open.
What the four workstreams do and what you receive
Most buyers start with one workstream and add another once the first has produced a measured result.
| Workstream | What the work involves | What you receive | Where a person decides |
|---|---|---|---|
| Vulnerability discovery | Models read code and configuration next to conventional scanners, and engineers reproduce each finding. Method: AI vulnerability discovery. | Reproduced findings, proposed fixes and retest evidence | Reproduction, fix approval and closure |
| SOC alert triage | Models summarize and enrich alerts and suggest a priority. Method: AI SOC alert triage. | Investigation summaries in the analyst queue and a running evaluation record | Every closure and every containment action |
| Incident response readiness | Playbooks, evidence handling and exercises written for incidents that involve AI systems. Method: AI incident response plan. | Tested playbooks and exercise findings | Declaring an incident and approving containment |
| Security for AI systems and agents | Threat models, tool-permission reviews and adversarial tests of LLM applications. Method: AI agent security checklist. | A threat model, reproducible test cases and control changes | Which permissions an agent receives and what ships |
The workstreams share evidence. A confirmed vulnerability can become a detection rule, and an incident exercise can expose a gap in alert handling.
Why prioritizing is harder than finding
Finding candidate weaknesses is no longer the scarce skill. Deciding which ones to fix first is, and public signals exist for that. CISA's Known Exploited Vulnerabilities catalog lists vulnerabilities known to be exploited in the wild and is meant as an input to prioritization.1 FIRST's EPSS publishes a daily probability, between zero and one, that a CVE will be exploited in the next 30 days.2 CVSS v4.0 keeps base severity separate from threat and environmental context.3
CISA's BOD 26-04, issued June 10, 2026, shows where practice is heading.4 It revoked BOD 22-01 and sets remediation timelines from four factors: public exposure, known exploitation, whether exploitation can be automated, and technical impact.4 The directive is addressed to federal agencies, but the same four questions are a sound starting point for any organization's patch policy.
A model can read a large codebase quickly and draft a hypothesis about a flaw. It cannot tell whether the flaw is reachable in your environment, who owns the system, or what a restart costs. Those answers come from engineers and from you, which is why every engagement ends in verified findings ranked by exploitability and business impact.
Choosing a fixed sprint or a managed service
Choose by how often the question comes back, not by budget.
- If
You need one answer by a set date, such as the exposure of an internet-facing estate before an audit or an acquisition.
ThenRun a fixed sprint with agreed deliverables.
A closed scope keeps the evidence deep enough to rely on, and the work stops when the deliverables are accepted.
- If
Alerts exceed analyst capacity and the same pressure returns every week.
ThenChoose a managed service for alert triage and enrichment, with your analysts signing off.
A recurring workload needs a recurring evaluation of how well the triage is working.
- If
You have just shipped an AI agent or a retrieval system to real users.
ThenStart with a sprint that threat-models the system and tests it adversarially.
Permissions and untrusted-content paths are cheaper to fix before wider rollout.
- If
Your own team already triages well and wants a second reviewer.
ThenAsk for a review of their process and evaluation data instead of a full service.
Paying for work your team has already done adds cost without adding certainty.
Fixed sprint compared with managed service
| Aspect | Fixed sprint | Managed service |
|---|---|---|
| Scope | One agreed question and a named list of systems | A standing scope reviewed on a set cadence |
| End point | The work ends when deliverables are accepted | The work continues until either side ends it, with notice |
| Output | A report, a prioritized backlog and retest evidence | Ongoing queue summaries, tuning records and periodic evaluation reports |
| Production changes | Proposed by the engagement team, approved by your named approver | Proposed by the engagement team, approved by your named approver and logged |
| Best for | Assessments, pre-launch reviews and one-time exposure questions | Repeating workloads such as triage and monitoring |
Both models keep the same rule: nothing changes in production without a recorded human approval.
What you should hold when an engagement ends
Where the standards fit
The work maps to published frameworks so that results can be read by auditors and boards. NIST's Cybersecurity Framework 2.0, released on February 26, 2024, has six functions: Govern, Identify, Protect, Detect, Respond and Recover.5 NIST SP 800-61 Revision 3, published in April 2025, applies incident response guidance as a CSF 2.0 community profile and replaced Revision 2.6
Detection and testing work is described in the vocabulary of MITRE ATT&CK, a knowledge base of adversary techniques based on real-world observations.7 Security work on AI systems is described against the OWASP Top 10 for LLM Applications, whose 2025 edition lists prompt injection, excessive agency and unbounded consumption among its ten risks.8
Model providers also publish access programs that give vetted defenders fewer cyber-related refusals. Their rules belong to the providers, and the pages describe them from the providers' own documents: see frontier AI models for cyber defense.
Frequently asked questions
What are AI cyber defense services?
They are security services in which language models do the reading-heavy work, such as reviewing code, summarizing alerts and drafting playbooks, while security engineers and analysts verify results and decide what changes. The scope here covers vulnerability discovery, SOC alert triage and enrichment, incident response readiness, and security for AI systems and agents.
Who can authorize testing, and what stays out of scope?
Only the owner of a system, or someone with written authority from the owner, can authorize testing. Shared platforms and third-party services need their own permission. Anything without a clear owner is recorded and excluded until someone with authority decides, so no engagement touches a system by assumption.
Does a model ever change production systems on its own?
No. Models propose, summarize and rank. A named person at the client approves every change to a production system, including containment actions such as isolating a host or disabling an account, and the approval is recorded with what changed and when.
What does a scoped exposure assessment cover?
It maps the internet-facing assets and services in the agreed scope, checks them for known exploited and likely exploitable weaknesses, and ranks results by exploitability and business impact. Every finding is reproduced by a person. The output is a prioritized backlog that decides whether a sprint or a managed service is the next step.
How do model providers' access programs for defenders matter?
Some providers verify defenders before reducing cyber-related refusals, and each sets its own eligibility, tiers and data conditions. Those terms apply to whichever organization applies and can change. The deep-dive page on this site summarizes what the providers state and links to their pages so you can check the current terms.
Can this replace an in-house security team?
No. The service supports a team and does not replace accountability for it. Your people keep the authority to close alerts, approve changes and declare incidents. What changes is the amount of reading, enrichment and first-pass analysis they must do before they decide.
Sources
- CISA: Known Exploited Vulnerabilities Catalog — CISA
- FIRST: Exploit Prediction Scoring System (EPSS) — FIRST
- FIRST: Common Vulnerability Scoring System version 4.0 — FIRST
- CISA: BOD 26-04, Prioritizing Security Updates Based on Risk — CISA
- NIST: NIST Releases Version 2.0 of Landmark Cybersecurity Framework — NIST
- NIST SP 800-61 Rev. 3: Incident Response Recommendations and Considerations for Cybersecurity Risk Management — NIST
- MITRE ATT&CK: Resources — MITRE
- OWASP: Top 10 for LLM Applications 2025 — OWASP
Capabilities
The expertise this service draws on
Insights · AI Cyber Defense
Articles from this practice
Practical analysis on the decisions this service helps with.

Vulnerability Prioritization When Exploitation Outpaces Patching
CISA's BOD 26-04 replaced fixed patch deadlines with a risk table. Learn how to combine KEV, EPSS, SSVC and CVSS v4.0 so triage keeps up as AI-assisted discovery raises finding volume.

Legacy Modernization With AI Coding Agents: Where They Help, Where They Fail
AI coding agents can speed up understanding, testing and migrating old systems, but only inside controls that prove the new code behaves like the old. Here is where they help, where they fail and what to put in place.

AI Vendor Risk Management Under DORA: Treat Model Providers as Critical Third Parties
A foundation model provider is an ICT third party under DORA, and UK rules add expectations on exit and concentration. See what to record, contract, test and plan before a supervisor asks.
More services
Other ways to work with us
- Operations & DeliveryManaged Services
- Advisory & ArchitectureStrategic Technology Consulting
- Engineering & SystemsCustom Software Development
- Deep Science ResearchFrontier R&D
- Finance & Token DesignCapital Raising & Tokenomics
- Talent & LeadershipExecutive Recruitment
- Global OperationsGeographic Expansion
- Applied AI AgentsUse Cases by Workflow
Your starting point
What would you like to move forward?
Start with the outcome that feels closest. We'll help give the next step a useful shape.
Let's find your next move
One conversation. A clearer direction.
Tell us what you want to improve. A few sentences are enough to start.
shayan@coldai.org
