
Capability
NewCyber Defense
We help organizations defend against attackers who now use AI, and secure the AI systems and agents they put into production. The work combines frontier models made available to vetted defenders with threat modeling, engineering controls and governance that keep a person accountable for every security decision.
What does ColdAI do in Cyber Defense?
An AI security strategy has two fronts that need one owner and one budget conversation: defending against attackers who use AI, and securing the models, assistants and agents your own organization deploys. This page sets out what leaders should decide, in what order, and what good looks like, with deeper guides on prompt injection, red teaming, threat reporting, regulation and supply chain.
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.
How we help
How does ColdAI help with Cyber Defense?
Attackers now use AI to write convincing phishing at scale, probe exposed services and turn published vulnerabilities into working exploits faster than most patch cycles. At the same time, every model, agent and retrieval pipeline an organization deploys opens a new attack surface: prompt injection, over-permissioned agents and model and tool supply chains nobody has vetted. ColdAI works on both fronts, using frontier AI to find and fix weaknesses in the systems you own, and engineering the controls that keep your own AI systems from becoming the way in.
- Exposure assessments that map what an attacker can reach, with every finding verified by a person
- AI-assisted vulnerability discovery in your code and infrastructure, with fixes verified in your environment
- Threat models for AI agents, tool permissions and retrieval pipelines before they reach production
- Red-teaming of AI systems against prompt injection, data leakage and unsafe tool use
- Security operations support that uses AI to triage and enrich alerts, with analysts signing off on every action
- Board and regulator reporting that ties cyber risk to business impact and current obligations
Offerings
What ColdAI offers in Cyber Defense
- 01
AI-Assisted Vulnerability Discovery
Code and infrastructure review using frontier models to find vulnerabilities at depth, with human verification of every finding and fixes retested before closure.
- 02
AI System & Agent Security
Threat modeling and engineering controls for LLM applications and agents: least-privilege tool access, input and output filtering, and isolation of untrusted content.
- 03
AI Red Teaming
Adversarial testing of models, agents and retrieval pipelines against the OWASP Top 10 for LLM Applications and MITRE ATLAS techniques, with reproducible test cases.
- 04
SOC Augmentation
AI-assisted alert triage, enrichment and investigation summaries that cut analyst workload while keeping containment and response decisions with your team.
- 05
Incident Response Readiness
Playbooks, tabletop exercises and evidence handling designed for incidents that involve AI systems, including model misuse, data exposure and compromised agents.
- 06
Cyber Governance & Board Reporting
Security programs mapped to the NIST Cybersecurity Framework 2.0 and sector rules, with reporting that boards and regulators can act on.
Our approach
How Cyber Defense engagements run
Every engagement starts from the systems you own and the decisions you need to make, and ends with fixes verified in your environment rather than a list of findings.
- Step 1
Scope & Authorize
Agree the systems in scope, the rules of engagement and who approves what. Testing only touches systems you own or are authorized to test, and production changes always need a named approver.
- Step 2
Discover & Verify
Use frontier models alongside conventional scanners to find exposed services, vulnerable code and weak AI controls, then have security engineers reproduce each finding so you only act on real issues.
- Step 3
Fix & Harden
Prioritize by exploitability and business impact, write or review the fixes with your engineers, and retest each one so every closed finding is proven closed.
- Step 4
Monitor & Improve
Feed what was learned into detection rules, incident response playbooks and AI system guardrails, and track exposure over time so the next assessment starts from a measured baseline.
In depth
Go deeper into AI security
Five working guides covering the controls, tests, reports, rules and supplier checks behind the two fronts.
- 01GuidePrompt injection defense: layered controlsDirect and indirect prompt injection, why filtering alone fails, and the layered controls that limit damage: privilege separation, isolation and monitoring.
- 02ProcessAI red teaming: a seven-step processRun AI red teaming from scoping to retest: threat model, attack library, execution, triage and reporting, mapped to OWASP LLM risks and MITRE ATLAS.
- 03Deep diveAI cyber threats: what official reports showWhat Anthropic, OpenAI, Google GTIG and the UK NCSC have published about AI cyber threats, where they agree, and what defenders should change first.
- 04Regulation explainerAI cybersecurity regulation in the EUHow Article 15 of the EU AI Act, the Cyber Resilience Act and DORA treat AI cybersecurity, with dates and status as checked in November, and what to prepare.
- 05ArchitectureAI supply chain security: a reference architectureSecure the AI supply chain: model signing with Sigstore, AI bills of materials, safer file formats than pickle, dataset risks and vetting MCP servers.
Two fronts that need one owner
Most conversations about AI and security blur two different problems. The first is external: attackers use AI to work faster and at larger scale against the systems you already run. The second is internal: your own models, assistants and agents create new ways to lose data or take a wrong action. Different teams usually own each problem, so gaps open between them.
On the first front, official reporting is consistent about what has changed. In its September 2026 misuse report, Anthropic said none of the operations it described relied on a wholly novel technique and that the change was in the economics1. The UK National Cyber Security Centre's 2026 annual review judges that AI is a force multiplier that mainly speeds up and scales existing attack methods rather than creating new attack vectors2. For a leader, the first front is mostly a question of tempo: patching, detection and response must move faster.
The second front is different in kind. OWASP's 2025 list of risks for LLM applications puts prompt injection first and also names supply chain, excessive agency and improper output handling3. The NCSC has warned that prompt injection may never be totally mitigated the way SQL injection can be, so the design has to limit what a manipulated system is able to do4.
Decisions for the first ninety days, in order
The order matters because each decision gives the next one something to work with.
Name an owner for each front
Assign one accountable executive for AI-enabled threats and one for the security of deployed AI, and have both report to the same forum. Handoffs between the two then have a name attached.
Inventory what you run
List every model, assistant, agent, tool connection and dataset in use, including tools staff adopted without approval. The NCSC expects unapproved AI tools used with company data to reduce visibility and increase exposure2.
Set what each agent may do
For every agent, write down the data it can read, the actions it can take and who approves the high-impact ones. OWASP's mitigation list for prompt injection includes least privilege and human approval for high-risk actions3.
Raise the tempo of patching and detection
The NCSC's 2025 assessment says the time between disclosure and exploitation has shrunk to days and that AI will almost certainly reduce it further5. Agree target times for fixing exposed, exploited vulnerabilities, then test whether detection can meet them.
Test before an attacker does
Run an authorized red team exercise against the highest-risk AI systems, with written rules of engagement and a retest. The guide on AI red teaming sets out the stages.
Prepare the response
Extend incident response to cover AI failures, such as an agent taking a wrong action or a model exposing data, and to meet reporting clocks. Reporting obligations under the EU Cyber Resilience Act began applying on 11 September 20266. See the AI incident response plan guide.
How the two fronts differ in practice
| Question | Front one: attackers using AI | Front two: your own AI systems |
|---|---|---|
| What changes | Speed and scale of familiar attacks such as reconnaissance, phishing and exploitation2. | New ways to lose data or trigger wrong actions through models, agents and the tools they call3. |
| Usual owner | CISO, security operations and vulnerability management. | Product, engineering and platform teams, with security review. |
| First control | Shorter patch and detection times for exposed systems. | Narrow permissions and approval gates for every agent, enforced outside the model. |
| Evidence to ask for | Time to remediate exploited vulnerabilities and results of detection tests. | Agent inventory, tool permission list and red team results. |
| Where regulation bites | ICT risk and incident duties for financial entities under DORA, which the European supervisory authorities say remain highly relevant to frontier AI risks7. | Cybersecurity requirements in Article 15 of the EU AI Act for high-risk AI systems8. |
Most organizations need both columns. The second column is where the time between a design decision and an incident is usually shortest.
What good looks like after a year
Use these as a self-assessment. Each item should be provable with a document, a test result or a log.
How regulatory dates shape priority
Regulation does not decide your security design, but it does set some deadlines. The European Commission's AI Act page says the rules for high-risk systems under Annex III apply from 2 December 2027 and the rules for systems embedded in regulated products from 2 August 2028, both extended by the AI Omnibus9. The Commission's Cyber Resilience Act page gives 11 December 2027 for the main obligations, after the reporting duties that began on 11 September 20266.
Financial entities have a separate signal. In a statement dated 31 July 2026, the European supervisory authorities encouraged financial entities to act fast and proactively, because shorter vulnerability discovery and exploitation cycles raise the risk7. Dates and texts change, so confirm current status with counsel; the regulation guide lists what was checked and when.
How the work is usually sequenced
A sound engagement runs in four steps: scope and authorize, discover and verify, fix and harden, then monitor and improve. The same sequence works whether the target is an exposed system that attackers could reach with AI assistance or an agent your own teams are about to release.
The first step is the most often skipped. Written authorization, a defined scope and agreed stop conditions protect both the organization and the people doing the testing. Everything after that depends on them.
Frequently asked questions
What is an AI security strategy?
It is the set of decisions that covers two problems together: defending against attackers who use AI to work faster and at larger scale, and securing the models, assistants and agents your own organization deploys. It names owners, sets what each AI system may do, sets patching and detection targets, schedules testing and prepares response and reporting. It is a management plan, not a tool purchase.
Do we need a separate program for securing AI agents?
You need controls built for how agents fail, but they should sit inside the existing security program with the same owners, reviews and reporting. Agents combine untrusted input, tool access and autonomy, so permissions, approval gates and logging need explicit design. Treat each agent like a new class of privileged service account, and review it when its tools change.
Is prompt injection something we can patch?
Not in the way SQL injection can be fixed. The NCSC says language models have no distinction between data and instructions, and that prompt injection may never be totally mitigated4. The practical response is to limit the privileges of any system that reads untrusted content, add deterministic safeguards outside the model, and monitor behavior closely.
Which AI security dates should a board know about?
For the EU, the Commission lists 2 December 2027 for Annex III high-risk AI rules and 2 August 2028 for high-risk systems embedded in regulated products9. The Cyber Resilience Act's reporting obligations have applied since 11 September 2026, and its main obligations apply from 11 December 20276. Check national law and counsel for anything specific to your sector.
Who should own AI security, the CISO or the AI team?
Accountability for outcomes should sit with one executive, usually the CISO, while the AI and engineering teams own implementing the controls. Without a shared forum, each side assumes the other has covered the gap. A joint review of the AI inventory, agent permissions and open red team findings each month keeps the responsibilities visible.
Should we block staff from using AI tools until security is settled?
Blocking alone tends to push use out of sight. The NCSC expects staff using unapproved AI tools with company data to reduce visibility over that data and increase exposure2. A better first step is to publish approved tools, give staff a route to request others, and log use where you can.
Sources
- Detecting and countering misuse of AI: September 2026 — Anthropic
- Annual Review 2026: the impact of AI on the cyber threat — UK National Cyber Security Centre
- LLM01:2025 Prompt Injection — OWASP GenAI Security Project
- Prompt injection is not SQL injection (it may be worse) — UK National Cyber Security Centre
- Impact of AI on cyber threat from now to 2027 — UK National Cyber Security Centre
- Safer and more secure digital products — European Commission
- ESA statement: toward a consistent and risk-based approach for ICT risks from frontier AI models (JC 2026 25) — European Supervisory Authorities
- Article 15: Accuracy, robustness and cybersecurity — European Commission, AI Act Service Desk
- AI Act: regulatory framework for AI — European Commission
Insights · Cyber Defense
Articles from this practice
Practical analysis of the decisions this expertise informs.

AI Agent Identity and Access Management: Treat Agents as Privileged Identities
AI agents with tool access can read, write and spend on their own. Here is how to give them distinct identities, least privilege, short-lived credentials, approval gates and audit trails.

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.
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