(01)
Hijacked agents
- The challenge
- An instruction planted in a document or web page can turn a helpful agent into an insider.
- Our work
- Open detection content for agent tool calls and a reference guardrail architecture.
AI-native cyber defence
Security engineering, AI SOC and AI deployment for organisations that cannot afford to guess.
- Security Engineering
- AI-Driven SOC
- LLM and Agent Security
- Secure AI Deployment
- Agentic AI Engineering
- AI Governance and Assurance
(Questions — 08)
Questions your board will ask.
- (01/08)
If one contractor laptop is phished tonight, how many hops until someone notices?
- (02/08)
Which log sources could you switch off tomorrow without losing a single detection?
- (03/08)
Your people already paste company data into public AI. Do you know which data?
- (04/08)
Can you prove what your AI agents did last week, and who approved it?
- (05/08)
What stops an instruction hidden in a PDF from making your assistant email the customer list?
- (06/08)
How long does a tier-1 alert wait before a person looks at it?
- (07/08)
The CFO’s voice calls finance at 6 p.m. asking for a wire. What happens next?
- (08/08)
Which AI models run in your estate today, and where did their weights come from?
We answer these for a living.
What we do.
(06 services)
(01)Security Engineering
Detection engineering, cloud and identity security, with coverage mapped to MITRE ATT&CK.
Explore(02)AI-Driven SOC
Private and hosted models built into SOC workflows to triage, investigate and cut noise, without data leaving your boundary.
Explore(03)LLM and Agent Security
Red teaming and runtime guardrails against prompt injection, tool abuse and data leakage, mapped to OWASP and MITRE ATLAS.
Explore(04)Secure AI Deployment
Private models, RAG and third-party AI connected to your systems with access control, audit trails and privacy enforced by design.
Explore(05)Agentic AI Engineering
Governed agents that do real work in security and operations, with scoped permissions, human approval and a full action log.
Explore(06)AI Governance and Assurance
ISO/IEC 42001 and NIST AI RMF programs, AI inventories and risk registers, and board-ready reporting.
ExploreNIST CSF 2.0NIST AI RMFISO/IEC 27001ISO/IEC 42001OWASP LLM Top 10OWASP Agentic AIMITRE ATT&CKMITRE ATLASIEC 62443
(01)TorontoHeadquarters
(02)06Services
(03)09Frameworks we map to
Problems solved.
(05 cases · client names withheld)
- (01/05)Financial services
Private AI reads the telemetry first
−62%Alert volumeRead the case : Private AI reads the telemetry first
Case study: Private AI reads the telemetry first (Financial services) - (02/05)Manufacturing
Third-party AI on a legacy production line
+6 ptsEquipment effectivenessRead the case : Third-party AI on a legacy production line
Case study: Third-party AI on a legacy production line (Manufacturing) - (03/05)Healthcare network
Agents on the night shift
6 minMedian triage, was 4 hoursRead the case : Agents on the night shift
Case study: Agents on the night shift (Healthcare network) - (04/05)Software company
Locking down agent tool access
41 → 0Unvetted MCP serversRead the case : Locking down agent tool access
Case study: Locking down agent tool access (Software company) - (05/05)Insurance
From shadow AI to certified AI
27Shadow AI tools retiredRead the case : From shadow AI to certified AI
Case study: From shadow AI to certified AI (Insurance)
The next five years of AI in security.
(04 challenges)
(02)
Machine-speed exploitation
- The challenge
- AI shortens disclosure-to-exploit from weeks to hours. Patch cycles cannot keep up.
- Our work
- Briefs within hours of new exploited CVEs, with KQL, SPL and Sigma detections.
(03)
Synthetic identity
- The challenge
- Cloned voices and live deepfakes make “call to verify” unreliable.
- Our work
- Verification playbooks and detections for MFA resets and unusual payments.
(04)
Poisoned AI supply chain
- The challenge
- Model weights and plugins arrive with little provenance and can run code on load.
- Our work
- Signed artefacts, an AI bill of materials and model-file scanning in every deployment.
(Briefing)
Start with one question.
Thirty minutes with an engineer. You leave with a clear answer, not a pitch.