Project OffSecAI — AI-Driven Security Auditing

Project OffSecAI is our internal R&D programme exploring how AI can amplify security auditing — and how to secure AI systems themselves. It operates as two complementary capabilities: AI-accelerated audit delivery for our standard penetration testing and cloud security work, and specialist AI security testing for organisations deploying LLM-based applications.

AI-Accelerated Audit Delivery

Traditional security audits spend disproportionate effort on enumeration, configuration analysis, and pattern matching — work that AI does faster and more comprehensively than humans alone. OffSecAI handles this phase, freeing human analysts to focus on what they do best: creative attack chaining, business-logic reasoning, and the judgement calls that distinguish a useful finding from a noisy one.

Every finding produced with OffSecAI assistance is reviewed and validated by a human analyst before reporting. AI without human review produces convincing-sounding nonsense. Human review without AI assistance produces narrower coverage in available time. The combination is materially more effective than either alone.

Practical effects:

• Faster audit cycles — typical penetration tests complete 30-40% faster with the same human team
• Broader coverage — configuration analysis at scale identifies findings traditional manual review misses
• Better reproducibility — automated test catalogues ensure consistent coverage across engagements
• Lower cost for SMEs — services previously priced only for enterprise become accessible to mid-market

Specialist LLM Application Security

The rapid adoption of AI in production systems has created a novel attack surface that traditional security testing does not address. We have developed specialist methodology for assessing LLM applications, agentic AI systems, and AI-integrated workflows.

The methodology aligns with the OWASP Top 10 for LLM Applications and MITRE ATLAS, and is documented in our LLM security article.

What We Test

• Direct prompt injection — instruction override and jailbreak techniques against system prompts
• Indirect injection — adversarial content embedded in documents, web pages, emails, and other data the AI processes
• Tool-call hijacking — manipulating agentic systems into invoking tools with attacker-controlled parameters
• Data exfiltration via outputs — covert channels in model responses, structured data leakage, URL-based exfiltration
• Context window poisoning — adversarial content designed to shift behaviour through volume or repetition
• Supply chain vulnerabilities — third-party SDK analysis, embedding model integrity, fine-tuning data provenance
• Excessive agency — identifying where AI systems have broader permissions than their stated tasks require
• RAG and embedding manipulation — injecting adversarial content into retrieval indices

Continuous Adversarial Testing

For organisations operating production agentic AI systems, point-in-time testing is insufficient. The threat landscape evolves faster than typical annual audit cycles. We offer continuous adversarial testing — monthly or quarterly assessments against an updated payload library — to detect regressions as system prompts, tool integrations, and model versions change.

Last quarter's clean bill of health does not survive contact with this quarter's threat techniques. Continuous testing is the only viable approach for high-stakes deployments.

Research Programme

The OffSecAI team publishes technical research on emerging AI security topics:

• LLM penetration testing benchmarks
• Prompt injection technique catalogues with mitigation analysis
• Vulnerability disclosure for AI platforms and integrations
• Threat modelling frameworks for agentic systems

Public research appears on our blog. Deeper technical material — including offensive tooling, payload libraries, and proof-of-concept exploit chains — is shared with active clients under NDA.

When to Engage OffSecAI

• Pre-launch security review of an AI-integrated product or feature
• Annual security testing of production LLM applications
• Post-architecture-change validation when system prompts, tool integrations, or model versions are updated
• Continuous monitoring for high-stakes agentic AI in production
• Board-level assurance review for AI risk reporting
• Acquisition due diligence on AI-heavy targets

UK & USA Coverage

OffSecAI engagements are conducted remotely from our UK and US facilities. Client data handling complies with UK GDPR for UK/EU clients and the relevant sector frameworks for US clients (HIPAA, GLBA, SOC 2). Where AI test payloads or proof-of-concept exploits raise dual-use export control considerations, we work with the client to define handling procedures appropriate to the jurisdiction.