?
InfluencePrompt injection, source confusion, malicious content, and adversarial instructions.
AI Security
AI changes what can see, decide, and act. Test the authority, not just the prompt.
Modern AI systems combine untrusted content, model behavior, retrieval, memory, tools, browser state, identities, and downstream actions. We test the whole trust chain.
Where influence becomes authority
?
InputPrompt · web · document◉
ModelInterpret · plan · generate◇
GatePolicy · identity · confirmation⚙
ToolsBrowser · API · retrieval▰
DataMemory · private context!
ActionRead · write · send · executeHow we test
Treat the model as one component in a larger security system.
The important failures often occur where model output crosses into tools, authenticated browser state, private retrieval, persistent memory, external services, or business actions.
AI attack surface
⌁
ContextSystem prompts, retrieval, memory, browser state, user identity, and hidden data.
◇
AuthorityTool permissions, approval gates, identity propagation, and action boundaries.
!
ConsequenceData exposure, unintended actions, cross-context influence, or control bypass.
Where depth goes
Test the places where AI becomes operational authority.
We focus on the boundaries that can turn persuasive text or model output into a security consequence.
Authority boundaries
?
Prompt & source trustDirect and indirect injection, provenance, instruction hierarchy, and context confusion.
⚙
Tools & agentsTool invocation, multi-step behavior, permissions, confirmation gates, and delegated actions.
▰
Retrieval & memoryCross-user exposure, data scoping, memory contamination, and sensitive context handling.
◎
Browser AIDOM context, frame boundaries, authenticated state, browser actions, and evidence capture.
Evidence over spectacle
Make the AI failure path repeatable.
AI behavior can vary. Useful findings need controlled inputs, captured context, observable actions, boundary conditions, and replay information that distinguish a security property from a one-off response.
Reproducible AI evidence
01
InputPreserve the influencing content and source.
02
ContextRecord relevant model, browser, identity, and tool state.
03
ActionCapture the observable security consequence.
04
ReplayRepeat under bounded conditions and compare controls.
Evidence before AI narrative.
Start with the authority boundary
Tell us what the AI can access, what it can act on, and what must never cross that line.
We will map the influence path, permission model, sensitive context, and evidence needed to test the system safely.
Define the trust boundary
?
InfluenceWhat can shape behavior▰
ContextWhat sensitive data exists⚙
ToolsWhat the system can invoke◇
AuthorityWhat actions are permitted✓
EvidenceWhat must be proven