Research
AI Security
AI creates new attack surfaces and gives attackers new tools. This area covers securing AI systems, the AI supply chain, deepfakes and AI-enabled fraud, and the use of AI in security operations.
Topics in AI Security
Securing AI Systems (prompt injection, data poisoning)
Defending models, prompts, data and agents from attack.
Deepfakes & AI-Enabled Fraud
Countering synthetic media and AI-assisted social engineering.
Overview
AI systems can be manipulated in ways traditional software cannot. Prompt injection can redirect a model through hidden instructions, poisoned data can corrupt what a model learns, and agents with broad permissions can be turned against the systems they serve. Resources such as the OWASP Top 10 for Large Language Model Applications and MITRE ATLAS catalog these threats.
At the same time, attackers use generative AI to produce convincing phishing, synthetic voices and deepfake video. Defenders are also adopting AI to triage alerts and investigate incidents faster.
What leaders need to get right
- Threat modeling. Include AI-specific attacks in security design reviews.
- Least privilege. Treat models and agents as identities with minimal access.
- Supply chain. Know where models, data sets and AI libraries come from.
- Verification. Update payment and approval processes so a convincing voice or video is never enough.
Research report
Agents are a new class of identity and attack surface
Governing Agentic AI 2026 covers agent identity, MCP and A2A connectivity, prompt injection, runtime limits, kill switches and incident response, with a tool and MCP server intake review.
Featured research report
Governing Agentic AI 2026
67-page playbook for governing enterprise AI agents: autonomy levels, risk tiers, agent identity, MCP controls and human oversight. Includes 10 ready-to-use templates.
Questions leaders should ask
- Have our AI applications been tested for prompt injection?
- Do we know the origin of every model and data set we use?
- Could a deepfake of an executive trigger a payment in our organization?
- Where is AI helping our security team, and how is it supervised?
Other research areas
Related research reports
In-depth reports with ready-to-use templates.
- Governing Agentic AI 2026$59.99 PDF report
- Third-Party AI Risk$59.99 PDF report
Related research reports
Reports connected to this topic.
Latest insights on AI Security
AWS Continuum Patches: Why a Pull Request Is Not Verified Remediation
AWS Continuum validates code findings and proposes fixes. An enterprise evidence funnel and six tests distinguish patch delivery from proven risk reduction.
Agent Sandboxes Are Not Permission Systems: Six Tests for Strands Box
AWS previewed Strands Box as an agent sandbox. Six enterprise tests distinguish sandbox restrictions from identity, approvals, data access, and recovery controls.
GitHub AI App Scanning: What Pull-Request Alerts Cannot Prove
GitHub's October 2026 AI application scanning preview adds useful PR alerts. Six tests distinguish source findings from runtime permissions and security.