Selected Technical Work — Project Summaries
Provable AI
Verification infrastructure for making AI-related execution evidence inspectable, auditable, and independently
verifiable.
Designed and implemented the policy, execution, evidence-generation, cryptographic verification, and
audit-validation pipeline.
- 593 passing automated tests
- Live verifier online
- Client-side cryptography
- No server-side trust required
- Policy is enforced before execution to prevent out-of-policy actions from executing.
- Evidence generation is an isolated pipeline stage separate from core execution.
- Verification is 100% client-side: reviewers can audit proofs with zero server access.
Public repository: futureaihub/provable-ai
Live verifier portal: zorynex.co
Limitations
- Verifies that recorded evidence is internally consistent and untampered — it does not verify that the AI's
underlying reasoning or output was correct.
- Trust still depends on the execution environment generating the trace honestly in the first place;
verification detects tampering after the fact, not a compromised recorder.
Agent Trust Benchmark
A deterministic verification framework that tests whether tool-using AI agents actually obey authorization
policies, perform the correct actions, and report outcomes truthfully — using observable evidence instead of
trusting agent explanations.
Designed and implemented the adversarial test harness, OPA/Rego policy evaluation layer, benchmark-owned state
observer, SHA-256 evidence chain recorder, and 11 deterministic evaluator checks.
- 11 deterministic checks
- 258 automated tests
- 30 red-team cases
- 18/18 real-agent replays with 100% hash chain match
- Validates: authorization enforcement, correct-target execution, role-escalation resistance,
tool-result/claim consistency, tamper-evident evidence integrity
Public repository:
futureaihub/agent-trust-benchmark
Limitations
- A focused verification harness, not a production authorization system or a broad claim of model safety.
- Evaluates agent behavior within instrumented benchmark scenarios and defined policy rules; novel tools
require corresponding test harnesses and state observers.
AgentGuard
Deterministic policy gateway evaluating every AI agent tool call before execution with replayable audit
trails.
Designed and implemented the deterministic policy gateway, rule evaluation layer, fail-closed behavior, audit
storage, replay mechanism, and test suite.
- Deterministic policy engine
- Fail-closed enforcement
- SQLite audit store
- Zero-model trust
Repository: futureaihub/agentguard (private) — interactive execution preview available on this page.
Limitations
- Does not attempt to prevent or detect prompt injection itself — it assumes the agent may already be
compromised and stops the resulting action instead.
- Policy coverage is only as good as the rules defined; an action type with no matching rule is blocked by
default, but that also means legitimate new tool calls need a rule added before they'll run.
Hybrid Website Scanner
Modular pipeline orchestrating ten security scanners, normalizing conflicting findings into a single
deduplicated risk report.
Designed and implemented the orchestration pipeline, scanner normalization, async job processing, correlation,
deduplication, and verification workflow.
- 10 tools orchestrated: Nuclei, OWASP ZAP, testssl.sh, WPScan, Subfinder, httpx, WhatWeb, Katana, Lighthouse,
axe-core
- Async job queue
- Docker containerized
- Deduplication engine
Repository: futureaihub/hybrid-website-scanner (private) — interactive execution preview available on this
page.
Limitations
- Orchestrates and correlates existing scanners — it does not implement novel vulnerability detection beyond
what the underlying ten tools already find.
- Findings still require human triage for business-logic-level risk; the pipeline reduces noise and
duplication, it does not replace judgment on severity in context.
AI Guard
Behavioral analysis inspecting LLM workflow inputs, outputs, and signals for injection and manipulation
patterns.
Designed and implemented the behavioral analysis layer, workflow context inspection, risk signal extractor, and
reviewable finding schema.
- Behavioral signals
- Prompt injection detection
- Context-aware analysis
- Reviewable output
Repository: futureaihub/ai-guard (private) — interactive execution preview available on this page.
Limitations
- Detects and flags suspicious behavioral signals — it does not guarantee prevention of every prompt injection
or manipulation attempt, and framing it as universal protection would overstate what it does.
- Ambiguous cases route to human review by design, which means throughput depends on reviewer capacity, not
just detection accuracy.