KNOWLEDGE ARTICLE / AI security · Defensive capabilities
Cyber security
against new AI threats.
AI-assisted attacks put pressure on the time available to find and fix weaknesses. This source-based overview explains what to improve in existing defences, how security capabilities can be built in stages and where AI can provide controlled support.
- Article date
- Base notes dated
- Latest source check
THE SHORT VERSION
Key points.
- Exposure and response time matter. An up-to-date inventory, faster remediation and working detection remain central to defence. [01] [02]
- Start with complete basic capabilities. Detection, a prepared response and tested recovery belong at the beginning. Later stages deepen them. [05]
- AI needs a defined task and limits. Analysis assistance can be introduced separately from authority to change systems. More autonomy is not a necessary goal. [15] [17]
01KNOWLEDGE ARTICLE
What changes for an existing defence
AI can help attackers find weaknesses, develop exploits and carry out attacks at greater speed and scale. The NCSC’s frontier-AI overview therefore places renewed emphasis on established controls. For organisations, the practical issue is how much time passes before an exposed weakness is removed and whether an intrusion is detected in the meantime. [03] [02]
The BSI warning of 22 June 2026 calls for patching exposed systems, particularly perimeter systems, within hours and preferably minutes. It also calls for stronger detection and an assessment of whether vulnerabilities have already been exploited. The timing concerns this exposed context. It is not a universal deadline for every system or approval for untested automatic updates. [01]
02KNOWLEDGE ARTICLE
Five priorities for improving existing protection
The following priorities bring together related guidance from the BSI and NCSC. They do not prescribe five new purchases. [01] [02]
Know and reduce external exposure
Keep systems, applications and their internet exposure visible, including services hosted in the cloud. Remove unnecessary access and restrict administrative interfaces. The BSI explicitly recommends external attack surface management (EASM), while NCSC prioritises the external attack surface when patching capacity is limited. [01] [04]
Connect findings to remediation
Assess actual exposure and business impact as well as technical severity. Automating advisory processing, for example with the Common Security Advisory Framework (CSAF), can support prioritisation. Before enabling fully automatic updates, assess their suitability and prepare backup and fallback mechanisms. Unsupported systems also need a plan for replacement or risk reduction. [01] [04]
Strengthen identities and contain the impact
Enforce multifactor authentication where supported, review privileges and separate systems and credentials that do not need to share access. Network segmentation limits the reach of a compromise. These principles also apply to service and AI-agent identities, whose tool access can otherwise connect an untrusted input to sensitive systems. [01] [17]
Operate detection and incident handling
Collect relevant events, protect logs outside the systems being monitored and ensure that alerts reach someone who can investigate and act. The BSI names SIEM and endpoint detection and response solutions among the relevant technologies. Where internal capacity is insufficient, it recommends considering external providers for incident response and secure operations. [01]
Test response and recovery
Maintain a usable response plan and protected backups, and test whether recovery works. NIST includes these tasks in its small-business entry guidance. Exercises should expose missing contacts, authority or recovery prerequisites before an incident creates time pressure. Detection and recovery therefore belong to the baseline, even if their technical implementation later becomes more sophisticated. [05] [02] [01]
03KNOWLEDGE ARTICLE
Prioritising safeguards with CIS
CIS assigns individual safeguards to Implementation Groups IG1, IG2 and IG3. Every organisation should start with IG1. IG2 adds safeguards to IG1; IG3 includes all safeguards. The appropriate group depends on the organisation’s risk profile and available resources. [06]
The matrix shows selected safeguards from CIS Controls v8.1. Their group assignments follow CIS; the selection and grouping into security areas serve this article. The columns do not prescribe a purchasing sequence. An urgent risk may require a safeguard shown further to the right at an early stage. [06] [01]
CIS IG1–IG3: comparing specific safeguards
Each column lists only newly added safeguards: IG2 also includes IG1, and IG3 also includes IG1 and IG2. Safeguard numbers identify the safeguards. “—” means no additional safeguard in this selection. The matrix does not replace a complete CIS assessment.
Assets, identities and data
| Security area | IG1Baseline | IG2Adds to IG1 | IG3Adds to IG1 + IG2 |
|---|---|---|---|
| Inventory |
|
|
|
| Identities and MFA |
|
|
|
| Data protection and DLP |
|
|
|
3.13 includes data held by service providers and does not promise complete prevention of data leakage. [08] | |||
Patching and preventive protection
| Security area | IG1Baseline | IG2Adds to IG1 | IG3Adds to IG1 + IG2 |
|---|---|---|---|
| Patches and vulnerabilities |
|
| — |
7.5 covers authenticated and unauthenticated scans. Routine intervals are not waiting periods for critical vulnerabilities. [07] | |||
| Endpoint and malware protection |
|
|
|
| Email and web |
|
|
|
9.2 includes devices outside the enterprise network. [23] | |||
Detection, response and recovery
| Security area | IG1Baseline | IG2Adds to IG1 | IG3Adds to IG1 + IG2 |
|---|---|---|---|
| Logs and attack detection |
|
|
|
| Incident response |
|
|
|
17.1 requires an internal oversight person even when incident response is outsourced. [26] | |||
| Backup and recovery |
|
| — |
Risk-based implementation and early recovery
Capabilities can be developed in parallel. In its municipal context, BSI WiBA also allows urgent improvements to be brought forward. Cloud services, applications and AI systems need protection from the start of their use. Critical and exposed systems need early attention; a higher IG is not a required target for every organisation. [11] [01] [06]
Detection, prepared response and tested recovery belong in early protection. NIST’s introductory guide includes recovery testing, whereas CIS assigns the explicit testing safeguard 11.5 to IG2. This difference is not a reason to postpone checking whether recovery works. [05] [25] [01]
NCSC: Foundation, Improver and Enhanced
The NCSC toolkit serves sole traders, micro businesses and small organisations. Its three layers prioritise actions by urgency and effort. Foundation is the recommended starting point, but actions may be completed in a different order. [10] [27]
The current toolkit assigns its main actions as follows: [28]
- Foundation: secure business email and important online accounts; manage passwords and passkeys; update devices and apps; secure devices.
- Improver: remove unnecessary user accounts; back up work files and data; separate administrator and user accounts; learn to recognise attacks; check antivirus and firewall protection.
- Enhanced: plan incident response; restrict cloud administration; review the digital footprint; consider Cyber Essentials.
These main actions have fixed layer assignments; individual steps adapt to information about the business. The layers do not correspond to CIS IG1–IG3: the toolkit places backups in Improver, while CIS includes automated backups in IG1. The toolkit therefore does not replace an enterprise-wide CIS implementation plan. [28] [10] [25]
04KNOWLEDGE ARTICLE
Other staged and prioritised approaches
The NSA also assigns concrete defensive measures to cumulative tiers. BSI WiBA offers a flexible implementation order for municipal administrations. CISA prioritises foundational security measures, while the NIST Tiers describe how risks are governed and managed. [29] [11] [30] [31]
NSA: from immediate protection to tested automation
The Best Practices Guide for Cyber Hygiene is aimed particularly at defenders, security teams and leaders in the US defence community. Its recommendations are more broadly applicable but assume basic cybersecurity knowledge. The NSA calls for Tier 0 first, followed by the enhancements in Tiers 1–3. The selection below follows the main text. [29]
Tier 0 already includes detection and response
Selected measures. Each tier builds on the previous ones. The three topic areas in Tier 0 are grouped for this article.
Tier 0 Reduce risk immediately
Inventory and maintenance
- Inventory hardware, software, accounts and data flows.
- Map assets to business functions.
- Validate automated patching and initiate configuration automation.
Restrict access and lateral movement
- Restrict accounts and privileges.
- Enforce phishing-resistant MFA universally.
- Segment critical systems and remove unnecessary network paths.
- Restrict application execution.
Detect and respond
- Validate SIEM log ingestion and refine EDR rules.
- Establish behavioural baselines and conduct continuous hunts.
- Validate response processes and security service providers’ response capabilities.
- Plan and exercise recovery.
Build on Tier 0
1 Prepare SOAR
- Plan use cases, test integrations and document processes.
- Complete logging and alerting in EDR and SIEM.
- Validate segmentation and eliminate shadow IT paths.
2 Implement and test SOAR
- Introduce SOAR and test responses in controlled environments.
- Connect SIEM, EDR and workflows.
- Document repeatable procedures.
- Involve people in AI-enabled processes where appropriate.
3 Emulate attacks
- Conduct red-team assessments.
- Analyse findings and have defenders remediate gaps.
- Exercise recovery again and maintain continuous threat hunting.
Segmentation, SIEM, EDR and exercised recovery therefore already belong to Tier 0. Higher tiers primarily deepen automation and validation. This model cannot be equated with CIS IG1–IG3 or the NIST Tiers. The NSA also explicitly states that the guide does not cover every attack vector and should not be the sole basis for hardening a network. [29] [06] [31]
CISA: prioritise measures by impact and effort
The voluntary CPGs describe a limited baseline of high-impact measures for IT and OT, particularly in critical infrastructure. Individual goals include cost, impact and ease-of-implementation ratings. They already cover responsibilities, security oversight, vulnerability remediation and incident preparedness. Their structure follows the six CSF Functions and does not prescribe cumulative implementation stages. [30]
NIST: the rigour of governance and risk management
The four CSF Tiers describe a progression from ad hoc responses to repeatable and continuously adapting practices. Leadership selects a target suited to risk, requirements and resources. The Tiers support comparison of current and target states, but do not assign individual security technologies to a purchasing sequence. [31]
05KNOWLEDGE ARTICLE
When an additional platform becomes useful
Existing endpoint and identity functions, patch management, protected backups and relevant logs can already provide essential capabilities. An additional platform needs a specific gap to address. Data sources, responsibilities and operating capacity determine whether it can be run effectively. [01] [05]
Further development extends these foundations: recurring vulnerability checks must lead to remediation and retesting, alerts need to be handled and coverage gaps understood. Depending on need, this may involve central analysis, Managed Detection and Response (MDR), deeper segmentation or further EASM functions. A provider can also support the initial implementation. [01] [05] [09]
Additional data, identity, network and application controls, specialised analysis and orchestration depend on protection needs. Automation requires dependable workflows and manageable consequences. This does not imply a mandatory package of additional platforms. [01] [06] [15]
EASM and vulnerability management
An inventory and awareness of external exposure are early requirements. Whether they need a separate EASM platform depends on the environment and existing tools. Vulnerability scanners likewise need a process that assigns findings, fixes them and checks the result. NCSC applies this prerequisite explicitly to AI-assisted vulnerability discovery. [01] [14]
Scan intervals and patch deadlines
CIS Control 7 (v8.1) distinguishes internal from externally exposed assets and specifies these minimum frequencies: [07]
- Internal assets: automated vulnerability scans at least quarterly, using both authenticated and unauthenticated scans (7.5, IG2 onwards).
- Externally exposed assets: automated vulnerability scans at least monthly (7.6, IG2 onwards).
- Operating systems and applications: updates through automated patch management at least monthly (7.3/7.4, IG1 onwards).
These routine intervals are not universal remediation deadlines from the moment a vulnerability becomes known. CIS specifies a risk-based remediation strategy. For exposed systems, particularly at the perimeter, the BSI warning instead calls for patching within hours and preferably minutes. The NCSC guidance considered here recommends faster, more frequent patching but gives no fixed scan intervals or numerical patch deadlines. [07] [01] [04] [02] [03]
EDR, XDR, SIEM and a managed service
Endpoint detection and response (EDR), extended detection and response (XDR) and security information and event management (SIEM) can cover different parts of detection and investigation. A product may bundle early and advanced functions. The BSI’s examples do not require all categories to be purchased separately. A managed service can provide operating capacity, while the choice still needs suitable data and clear responsibilities. [01] [09] [05]
DLP and automated response
For data loss prevention (DLP), sensitive data and relevant access or transfer paths must be understood. A formal IG3 assignment does not make data protection a late task. Security orchestration, automation and response (SOAR) is named by the BSI as an additional option for automated response. Its value depends on which procedures it can safely execute. Conventional rule-based automation does not require generative AI. [08] [01] [15]
06KNOWLEDGE ARTICLE
Use AI for a defined part of the work
The BSI lists checking infrastructure for weaknesses, updating inventories and handling routine work as supporting AI tasks. For software manufacturers it adds vulnerability discovery in code, testing products and patches, and evaluating detection rules or workarounds before a fix is available. These are supplementary uses. The warning does not recommend a specific vendor product. [01]
The NSA makes its defensive AI recommendation concrete in Appendix A of the cyber hygiene guide. AI should analyse inventory and log data to detect anomalies, suggest possible vulnerabilities and prioritise risks. Its examples include software with supply-chain vulnerabilities, suspicious connections to algorithmically generated domains and patterns of privilege abuse. It requires human validation of the findings. As best practice, it recommends combining AI with human threat hunting and reassessing after mitigation. [29]
The NSA calls for complete datasets as the basis for analysis. Their processing needs to meet protection requirements, particularly for inventories, logs and account information. The recommendation does not specify a generative model or a separate AI product. For automated responses, the guide describes previously tested SOAR integration and human involvement where appropriate. An AI finding therefore does not provide blanket authorisation to change production systems. [29] [01] [17]
NCSC recommends explainable assistance to people as an accessible starting point. An assistant can help prepare an assessment without being authorised to change production systems. Findings still need evidence and sufficient capacity for remediation. A faster summary or a larger number of findings alone does not demonstrate better protection. [15] [14]
Separate analysis from authority to act
An editorial comparison of operating modes, not a required progression towards autonomy. Permissions and action checks must be enforced outside the model.
Analysis assistance
- Task
- Prepare findings, context or a proposed next step from identified evidence.
- Required control
- Validate claims against original evidence and account for sensitive data and untrusted inputs.
A checked response workflow
- Task
- Use a finding to prepare or trigger a defined response procedure.
- Required control
- Test the procedure, log execution and provide a fallback for mistakes or operational disruption.
A narrowly scoped agent
- Task
- Choose multiple investigation steps and tools within a defined task.
- Required control
- Evaluate impact and recoverability, monitor activity and retain a way to stop or revoke access.
07KNOWLEDGE ARTICLE
Defensive AI also needs protection
An investigation assistant may process emails, logs, tickets or web content influenced by an attacker. An instruction hidden in that material can redirect the model. NCSC explains why prompt injection cannot simply be treated like an input-validation problem with a reliable separation between data and instructions. A system prompt alone cannot enforce a security boundary. [16]
The surrounding system must therefore enforce permissions, isolate access and check consequential actions. The BSI calls for isolation, validation layers and permission concepts for AI components. Human review needs usable evidence, and the consequences of an incorrect action must remain manageable. Adding a reviewer does not remove the need to constrain the agent’s technical access. [01] [17] [15]
08KNOWLEDGE ARTICLE
Manage cyber hygiene through an ISMS
Cyber hygiene also includes organisational measures. The CISA CPGs already require clear responsibilities, reviewed security policies and exercised response plans. This work can become part of an information security management system (ISMS). Whether a catalogue of measures is sufficient for an organisation still needs to be assessed against its risks and requirements. [30] [32]
An ISMS defines scope, security objectives and decision-making authority. Risks are assessed and treated, with residual risks accepted by the responsible management. Accountable people review implementation and effectiveness and initiate improvements. This also makes it possible to trace when a higher level of protection or an additional measure is needed. [32]
For the guidance compared here, this means incorporating existing measures and workflows into that management process. CIS or NSA measures do not need to be implemented a second time. The scope and documentation of an ISMS can be adapted to organisational size and needs. A management system standard can also be implemented without certification. Any additional requirement for formal assurance needs separate consideration. [32] [33] [34]
09KNOWLEDGE ARTICLE
Distinguish research questions from deployment guidance
The NCSC’s October 2026 AI chapter addresses redesigned workflows, AI-assisted attack-surface work, detection and the evaluation of autonomous defence. Its introduction and problem book identify unresolved research questions. They do not demonstrate that available products can already perform those tasks reliably or replace an organisation’s security function. [12] [13]
For an adoption decision, published use cases therefore need to be combined with checks in the intended environment. The problem book identifies comparison with human-led approaches as a research need. NCSC’s guidance on vulnerability discovery also addresses practical experience, result validation and data and operating conditions. The relevant question is whether a defined task is handled more reliably or effectively within acceptable error consequences. [13] [14] [15]
AI1–AI7: open questions on AI and cyber security
The seven topics range from AI-assisted defence and the security of AI systems to national resilience. This overview summarises selected sub-problems where the NCSC identifies a need for research. [13]
AI1 · Workflows: The NCSC asks how security processes can be redesigned for agents. This includes task allocation, tool access and meaningful human oversight, as well as the benefits and risks of multiple agents working together. [35]
AI2 · Attack surface: AI could combine inventories, vulnerabilities and operational dependencies to prioritise risks and suggest hardening measures. Open questions include handling incomplete data and developing transparent metrics for actual risk reduction. [36]
AI3 · Detection and response: Research covers combining weak signals, prioritising alerts, monitoring supply chains and maintaining a shared incident picture. AI recommendations should communicate uncertainty and support accountable human decisions. [37]
AI4 · Effectiveness and reliability: Research is needed on realistic testing, comparisons with human-only and human-AI teams, and ongoing assessment during operation. This includes failures, misuse and the loss of effective control. [38]
AI5 · Individual AI system lifecycles: Security questions span training data, fine-tuning and model weights, deployment and operation, through to retirement. Further research questions concern changes in deployed models, trade-offs involved in security controls and systems with constrained resources. [39]
AI6 · AI ecosystems: The focus is on dependencies between models, data, libraries, interfaces and agents. Open questions concern component provenance, AI-specific vulnerability management and the propagation of errors or attacks across connected systems. [40]
AI7 · National resilience: Knowledge graphs and other data-driven methods could reveal risks across organisations and sectors. Research is needed on data quality, provenance, uncertainty and reusable analysis. The focus is at national scale. [41]
Conclusion
Further improvements can address unknown exposure, slow remediation, unhandled alerts or recovery that has not been tested. AI can support work on these gaps, but it needs usable evidence, a responsible operator and controlled access. A staged expansion remains risk-based, including when an urgent need brings a capability forward. [01] [05] [14] [17]
REFERENCES / 41
Sources.
This article draws on public sources checked on 25 October 2026. The CIS matrix uses the official safeguard assignments; the selection and summaries of measures are specific to this article. NCSC layers, other operational guidance and open research questions are considered separately. The article documents neither an operational rollout nor a product test or a measured security improvement.
- BSI-IT-Sicherheitsinformation 2026-262788-1032 (opens in a new tab)
Version 1.0, pp. 4–7. The URL parameter v=2 does not change the version printed in the document.
- The AI shift in cyber risk: why leaders must act now (opens in a new tab)
Leadership priorities, exposure, patching and incident exercises.
- Frontier AI: what you need to know (opens in a new tab)
Overview of threat changes and opportunities for defence.
- Preparing for a vulnerability patch wave (opens in a new tab)
Prioritising external exposure and preparing patch capacity.
- NIST SP 1300: The NIST Cybersecurity Framework (CSF) 2.0: Small Business Quick-Start Guide (opens in a new tab)
PDF pp. 2, 5–8. Entry guidance includes detection, response and tested backups.
- CIS Critical Security Controls Implementation Groups (opens in a new tab)
IG1 as the starting point, with further safeguards selected by risk and resources.
- CIS Control 7: Continuous Vulnerability Management (opens in a new tab)
Assessment Specification v8.1, safeguards 7.2–7.7: risk-based remediation, automated patching, and different scan intervals for internal and externally exposed assets.
- CIS Control 3: Data Protection (opens in a new tab)
Assessment Specification v8.1, safeguards 3.2, 3.3, 3.6, 3.7, 3.10, 3.11, 3.13, 3.14.
- CIS Control 13: Network Monitoring and Defense (opens in a new tab)
Assessment Specification v8.1, particularly 13.1 centralised alerting, 13.7 host-based intrusion prevention and 13.11 alert thresholds.
- Small businesses receive cyber security boost with new toolkit from experts (opens in a new tab)
Foundation, Improver and Enhanced for sole traders, microbusinesses and small organisations.
- WiBA: Empfohlene Bearbeitungsreihenfolge (opens in a new tab)
pp. 3–4. Municipal-government context, flexible priorities and parallel work.
- Introducing a new AI chapter in the NCSC research problem book (opens in a new tab)
Introduction to a research agenda, not a product evaluation.
- AI in cyber security problems (opens in a new tab)
Chapter overview and framing of the research agenda. The seven individual topics are also supported by their respective subpages.
- 10 questions to ask when using AI models to find vulnerabilities (opens in a new tab)
Prerequisites, validation, data conditions and capacity to remediate findings.
- One does not simply defend agentically (opens in a new tab)
Human-facing assistance and risk dimensions for agentic defence. The dimensions are not maturity stages.
- Prompt injection is not SQL injection (it may be worse) (opens in a new tab)
Limits of treating untrusted language as safely separated data.
- Careful adoption of agentic AI services (opens in a new tab)
Joint guidance on cautious adoption, scoped access and low-risk, non-sensitive uses.
- CIS Control 1: Inventory and Control of Enterprise Assets (opens in a new tab)
Assessment Specification v8.1, safeguards 1.1, 1.3, 1.5; measures and Implementation Groups.
- CIS Control 2: Inventory and Control of Software Assets (opens in a new tab)
Assessment Specification v8.1, safeguards 2.1, 2.4; measures and Implementation Groups.
- CIS Control 5: Account Management (opens in a new tab)
Assessment Specification v8.1, safeguards 5.1, 5.6; measures and Implementation Groups.
- CIS Control 6: Access Control Management (opens in a new tab)
Assessment Specification v8.1, safeguards 6.3–6.8; measures and Implementation Groups.
- CIS Control 8: Audit Log Management (opens in a new tab)
Assessment Specification v8.1, safeguards 8.1–8.3, 8.9, 8.11, 8.12; measures and Implementation Groups.
- CIS Control 9: Email and Web Browser Protections (opens in a new tab)
Assessment Specification v8.1, safeguards 9.1–9.3, 9.5, 9.7; measures and Implementation Groups.
- CIS Control 10: Malware Defenses (opens in a new tab)
Assessment Specification v8.1, safeguards 10.1, 10.2, 10.6, 10.7; measures and Implementation Groups.
- CIS Control 11: Data Recovery (opens in a new tab)
Assessment Specification v8.1, safeguards 11.1–11.5; measures and Implementation Groups.
- CIS Control 17: Incident Response Management (opens in a new tab)
Assessment Specification v8.1, safeguards 17.1–17.4, 17.7–17.9; measures and Implementation Groups.
- NCSC Cyber Action Toolkit: How it works (opens in a new tab)
What are layers of protection? / What are actions? Prioritisation and flexible action order.
- NCSC Cyber Action Toolkit: public action configuration (opens in a new tab)
Publicly served content configuration retrieved on 2026-10-25: q.layers and the main actions for Foundation (6), Improver (5), Enhanced (4). Direct retrieval; no personalised profile was completed.
- Best Practices Guide for Cyber Hygiene (opens in a new tab)
Version 1.0, September 2026. Main text pp. 2–5 and Appendix A, p. 6. Analysis of the supplied full text; direct PDF retrieval unavailable. Tier assignments follow the main text.
- Cybersecurity Performance Goals 2.0 (CPG 2.0) (opens in a new tab)
CPG 2.0: introduction; Govern 1.A–1.C; Identify 2.B–2.C; cost, impact and ease of implementation; limited-scope statement.
- NIST SP 1302: Quick-Start Guide for Using the CSF Tiers (opens in a new tab)
NIST SP 1302, PDF pp. 2–3. Selection and use of Tiers for governance and risk management.
- BSI-Standard 200-1: Managementsysteme für Informationssicherheit (ISMS) (opens in a new tab)
BSI Standard 200-1, pp. 6, 29–31 and 34. Scope, risk decisions, accountability, effectiveness and proportionate implementation.
- ISO/IEC 27001:2022 — Information security management systems (opens in a new tab)
What is ISO/IEC 27001? / Who needs ISO/IEC 27001? ISMS, risk management and adaptation to size and needs.
- Management system standards (opens in a new tab)
What is a management system? / Certification. Appropriate complexity and implementation without certification.
- AI1 - How can organisations re-design cyber defence workflows to maximise the benefits of agentic AI? (opens in a new tab)
Individual research question, particularly “Strands or sub-problems”. Selected aspects summarised, not a complete catalogue of controls.
- AI2 - How can AI agents reliably help defenders identify, prioritise and reduce their attack surface? (opens in a new tab)
Individual research question, particularly “Strands or sub-problems”. Selected aspects summarised, not a complete catalogue of controls.
- AI3 - How can AI strengthen threat detection, triage and incident response? (opens in a new tab)
Individual research question, particularly “Strands or sub-problems”. Selected aspects summarised, not a complete catalogue of controls.
- AI4 - How should autonomous or semi-autonomous cyber defence systems be evaluated and assured? (opens in a new tab)
Individual research question, particularly “Strands or sub-problems”. Selected aspects summarised, not a complete catalogue of controls.
- AI5 - How can individual AI and machine learning systems be secured throughout their lifecycle? (opens in a new tab)
Individual research question, particularly “Strands or sub-problems”. Selected aspects summarised, not a complete catalogue of controls.
- AI6 - How do we secure current and future AI ecosystems? (opens in a new tab)
Individual research question, particularly “Strands or sub-problems”. Selected aspects summarised, not a complete catalogue of controls.
- AI7 - How can data-driven methods improve cyber security analysis, posture and resilience at national scale? (opens in a new tab)
Individual research question, particularly “Strands or sub-problems”. Selected aspects summarised, not a complete catalogue of controls.