Skip to content
View Pyhroff's full-sized avatar

Highlights

  • Pro

Organizations

@zka19-team

Block or report Pyhroff

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Pyhroff/README.md

Praising Harris Ratnam

AI Systems Security Researcher · Runtime Defense & Quantum Systems Engineer

GitHub LinkedIn Email


Upstream Open-Source Contributions & Formal Verification Badges

SQLGlot Merged LiteLLM Merged LiteLLM BFLA Fix LiteLLM Budget Auth PennyLane Open Google Qualtran Open FastMCP Open SQLFluff Approved PyPI Lean 4 Proofs Z3 SMT TryHackMe


Executive Summary & Core Engineering Philosophy

I engineer deterministic runtime defenses, formal protocol verification harnesses, and optimizing quantum/AST compilers for mission-critical software systems. My core thesis rejects unverified probabilistic safety assertions: probabilistic systems cannot reliably defend against adaptive adversaries without strict mathematical invariants at runtime boundaries.

flowchart LR
    A["Threat / Physical Model"] --> B["Formal SMT / Spectral Analysis"]
    B --> C["Empirical Micro-Benchmark Harness"]
    C --> D["Regression CI / CD Gate"]
    D --> E["Upstream Production Patch"]
    
    style A fill:#0F172A,stroke:#38BDF8,stroke-width:2px,color:#F8FAFC
    style B fill:#0F172A,stroke:#10B981,stroke-width:2px,color:#F8FAFC
    style C fill:#0F172A,stroke:#F59E0B,stroke-width:2px,color:#F8FAFC
    style D fill:#0F172A,stroke:#8B5CF6,stroke-width:2px,color:#F8FAFC
    style E fill:#0F766E,stroke:#34D399,stroke-width:2px,color:#FFFFFF
Loading
  • Zero-Trust Host Boundaries: Treat all language model completions, tool schemas, and third-party remote calls as untrusted inputs. Defenses reside in host compilers, kernel syscall monitors, and SMT-verified state machines.
  • Empirical Ground Truth: Every benchmark metric in this portfolio is backed by reproducible automated suites, SQLite databases, or machine-checked Lean 4 certificates. Zero speculation, zero fabricated claims.

🏛️ Executive Portfolio Highlights

┌───────────────────────────────────────────────────────────────────────────────────────────────────────┐
│                                 EXECUTIVE PORTFOLIO AT A GLANCE                                       │
├──────────────────────────────────────┬────────────────────────────────────────────────────────────────┤
│ 1. Upstream AI Gateway Security     │ LiteLLM PR #47312 (BFLA), PR #47311 (Budget Auth), Lean 4/Z3   │
│ 2. Fault-Tolerant Quantum Compilers  │ PennyLane PR #10391, Qualtran PR #1986, Q-MIND (58.5% T-Gates) │
│ 3. AI Security Suite & ModelHawk     │ aisec-suite (5 Planes, 0.03% FPR), ModelHawk (Pickle AST RCE) │
│ 4. Automated Adversarial Red Teaming │ PromptStrike: PAIR, TAP, Crescendo, GCG + CI Regression Gate   │
│ 5. Cyber Threat Intelligence (CTI)   │ DarkDecoder: MITRE ATT&CK + ATLAS Kill-Chain Reconstruction    │
└──────────────────────────────────────┴────────────────────────────────────────────────────────────────┘

Highlight 1: Upstream AI Infrastructure & Gateway Security (LiteLLM, FastMCP, SQLGlot)

BerriAI/LiteLLM: Multi-Tenant Access Control, FinOps Enforcement & Formal Invariants

  • PR #47312 (Broken Function Level Authorization - BFLA Fix): Identified and resolved a critical authorization flaw where tenant API keys could invoke /v1/mcp/access_groups to enumerate and bind cross-organizational MCP servers without tenancy checks (self-assessed CVSS 7.5). Patched by enforcing organization-scoped tenancy checks before returning server descriptors.
  • PR #47311 (Proxy Key Budget Enforcement): Fixed spend-limit tampering vulnerability in /key/update by requiring explicit authorization checks on budget_duration and budget_id modifications, preventing unauthorized callers from resetting or inflating financial usage caps.
  • PR #47105 (Broken Object-Level Authorization - BOLA Bypass): Fixed authorization flaw in can_object_call_model() where keys and teams with access-group IDs but an empty models: [] list defaulted to an unrestricted wildcard permitting access to all models.
  • PR #47091 (Merged by BerriAI): Standardized Cohere string stop parameter mapping to a list of strings (stop_sequences) in CohereChatConfig.
  • Formal Verification via Lean 4 & Z3 SMT (Commit 846d37e8): Identified fail-open MCP tool resolution on unvetted URLs and non-hyphenated dictionary inputs. Synthesized CEGIS guard conditions, verified inductive safety $\forall k \in [0, \infty)$ using Z3 SMT, and compiled machine-checked proof certificates in Lean 4 (lean_certs/MCP_04.lean).
┌──────────────────────────────────────────────────────────────────────────────────────────────────────┐
│                  LiteLLM PROXY ACCESS CONTROL & BUDGET SECURITY ARCHITECTURE                         │
├──────────────────────────────────────────────────────────────────────────────────────────────────────┤
│ Client Request ──► [Bearer API Key Auth]                                                             │
│                         │                                                                            │
│                         ▼                                                                            │
│             [Tenant Scope Verification] ◄── PR #47312: Enforce Org Boundaries on /v1/mcp/access_groups│
│                         │                                                                            │
│                         ▼                                                                            │
│             [Model Authorization Gate]  ◄── PR #47105: Block Wildcard Fallback on Empty models: []   │
│                         │                                                                            │
│                         ▼                                                                            │
│             [Spend & Budget Guard]      ◄── PR #47311: Enforce Authorization on /key/update Budgets  │
│                         │                                                                            │
│                         ▼                                                                            │
│             [SMT-Verified Dispatch]     ◄── Commit 846d37e8: Lean 4 / Z3 Verified Safe Tool Routing  │
└──────────────────────────────────────────────────────────────────────────────────────────────────────┘

PrefectHQ/FastMCP: Guest MicroVM Starvation Defense (PR #6063)

  • Vulnerability Remediated: In FastMCP's MontySandboxProvider (CodeMode), guest Python code instruction counts were capped inside the microVM, but external host callbacks escaped CPU counters, enabling CWE-653 / CWE-400 host event-loop starvation.
  • Mechanism: Implemented wall_clock_timeout_secs and max_external_calls callback budget caps with fail-closed cancellation. 65/65 tests pass cleanly.

Toby Mao/SQLGlot: Transpiler AST Mutation Bugfix (PR #8780 - MERGED)

  • AST Bug Resolved: In sqlglot/parser.py::_replace_lambda, transpiling higher-order functions with multi-argument lambdas mutated shared AST DataType nodes by reference, corrupting downstream dialect optimizers.
  • Mechanism: Replaced shared reference with .copy() on the AST node (exp.Cast(this=arg, to=data_type.copy())), passing all 9 dialect test suites without breaking AST invariants.

Highlight 2: Quantum Compilation & Fault-Tolerant Systems (PennyLane, Qualtran, Q-MIND)

                                  +------------------------------------+
                                  | Quantum Circuit Bloq Graph / AST   |
                                  +-----------------+------------------+
                                                    |
                                                    v
                     +--------------------------------------------------------------+
                     |                 Q-MIND Kickmix Optimization Engine            |
                     |  - Gidney Measurement-Based Uncomputation (4T / 0T uncompute)|
                     |  - Low-Chuang SelectSwap QROM Table Lookup (lambda = 2^k)    |
                     |  - Commutative Phase-Polynomial T-Gate Fusion                |
                     +------------------------------+-------------------------------+
                                                    |
                         +--------------------------+--------------------------+
                         |                                                     |
                         v                                                     v
      +--------------------------------------+             +--------------------------------------+
      | Google Qualtran (PR #1986 Open)      |             | Xanadu PennyLane (PR #10391 Open)    |
      | Cuccaro Adder CNOT & T Resource Fix  |             | Hamiltonian PyTree Autodiff Binding  |
      +--------------------------------------+             +--------------------------------------+
                         |                                                     |
                         +--------------------------+--------------------------+
                                                    |
                                                    v
                     +--------------------------------------------------------------+
                     |    64-Bit SIMD Kickmix Equivalence Fuzz Verifier (486M st/s) |
                     |      0 Bit Mismatches · 0 Ancilla Leaks · 100.0% Pass Rate   |
                     +--------------------------------------------------------------+
  • Xanadu PennyLane (PR #10391): Registered Hamiltonian Evolution operator PyTree leaves and fixed bind_new_parameters dispatch, resolving fatal JIT autodiff parameter-shift tracing crashes during quantum variational optimization (#6514).
  • Google Qualtran (PR #1986): Corrected CNOT and T-gate estimation formulas in Cuccaro adder Add.build_call_graph for asymmetric register bitwidths ($a \ne b$), restoring accurate fault-tolerant resource estimation for Shor's and quantum chemistry simulation algorithms (#1818).
  • Q-MIND Autonomous Quantum Compiler (Pyhroff/Q-MIND):
    • T-Gate & Toffoli Reduction: Achieved 58.53% T-gate reduction, 48.73% Toffoli reduction, and 21.37% depth reduction across Fusion Energy ($1/\sqrt{x}$ stopping power) and Decoded Quantum Interferometry ($\mathrm{GF}(2^m)$ Berlekamp-Massey decoder).
    • Equivalence Verification: Engineered a 64-bit SIMD bit-parallel Kickmix equivalence fuzz verifier executing at 486M gate-states/s with 100.0% pass rate (0 bit mismatches, 0 ancilla leaks across all trials).
    • Riemannian QNG: Pullback Fubini-Study metric tensor with adaptive Tikhonov regularization yielded a 5.3× convergence speedup vs Adam and 51.1% mean loss reduction.
    • Empirical Scale: Evaluated over 5,060 verified trials across 20 SQLite databases with 27 300-DPI publication figures and 172 passing tests.

Highlight 3: AI Security Suite & ModelHawk (aisec-suite, ModelHawk)

aisec-suite: Multi-Plane AI Security Umbrella & Runtime Scanner

Published on PyPI as aisec-suite v0.3.0 · 12 Modules across 5 Mathematical Security Planes · SARIF 2.1.0

┌────────────────────────────────────────────────────────────────────────────────────────┐
│                        AISEC-SUITE: 5-PLANE SECURITY ARCHITECTURE                      │
├──────────────────────────────────────────┬─────────────────────────────────────────────┤
│ Plane 1: Formal Protocol Verification    │ Plane 2: Weight Spectral Forensics          │
│ specprove-sec (Z3 SMT + 13 Lean 4 Certs) │ weightspecter (RMT Marchenko-Pastur Law)    │
│ - Bounded Model Checking (k-induction)   │ - BBP outlier spikes (350x - 976x detected) │
│ - CEGIS automated guard synthesis        │ - Surgical rank-1 deflation (0% FPR)        │
├──────────────────────────────────────────┼─────────────────────────────────────────────┤
│ Plane 3: Verifiable Inference & Checks   │ Plane 4: Activation Refusal Firewall        │
│ attestinfer (Freivalds Matmul P <= 2^-16)│ circuitshield (Residual Stream Clamping)    │
│ - 16-round Rademacher randomized checks  │ - Ortho-ablation & CAA steering mitigation  │
│ - SHA-256 Merkle activation audit tree   │ - Preserves refusal direction integrity     │
├──────────────────────────────────────────┴─────────────────────────────────────────────┤
│ Plane 5: Encrypted Network Side-Channel Defense & Traffic Shaping                      │
│ tokenspecter (Bayesian Viterbi DP + Ring-Buffer Bucket Pacer)                          │
│ - Eliminates SSE token timing leakage: Mutual Information 1.985 -> 0.000 bits/token     │
└────────────────────────────────────────────────────────────────────────────────────────┘
  • Production MCP Benchmark: Evaluated against 3,974 real-world production MCP tools: achieved 48.1% held-out recall at a 0.03% false-positive rate (only 1 false positive across 3,974 clean tools).
  • Labelled Triage: Evaluated on 120 labelled vulnerability findings with zero wrong refutations and deterministic exit codes (0 clean, 1 findings detected, 2 execution crash).
  • AgentDojo Defense: Runtime provenance guard halts 56.9% of AgentDojo attacks (99.3% banking, 95.2% Slack) without reading prose intent.
  • Supply-Chain & Memory Analysis: Integrates memsentry (memory safety analyzers) and mcpaudit with SARIF 2.1.0 output for GitHub Security Center integration.

ModelHawk: Static AST Disassembler for Pickle Deserialization Exploits

Pure Python Standard Library · Zero AI Dependencies · Zero Weight Execution Risk

Untrusted Model Weight File (.pt / .bin / .pth)
   │
   ▼
[Pickle Stream Tokenizer (pickletools)]
   │
   ├── Disassemble Opcode Instruction Stream (REDUCE, GLOBAL, BUILD, INST, OBJ)
   ├── Extract Target Callables: e.g. posix.system, subprocess.Popen, builtins.eval
   └── Trace __reduce__ Argument Construction Tuple
   │
   ▼
[Security Invariant Policy Engine]
   ├── SAFE: Pure PyTorch Tensor Metadata (torch.storage, collections.OrderedDict)
   └── DANGER: Disallowed Callable or Arbitrary Syscall Execution ──► HALT & QUARANTINE
  • Mechanism: Inspects PyTorch model weights by disassembling the opcode stream directly via Python's standard library pickletools. Detects code execution chains (__reduce__, GLOBAL, REDUCE) before pickle.load() is ever invoked, preventing host compromise.
  • Verification: 8/8 tests pass offline; includes benign and weaponized model generators for security regression testing. Technical analysis published on Medium: "I Built a Malware Scanner for AI Models. Here's What I Found".

Highlight 4: PromptStrike Automated Red Teaming & Defense Shield

Automated Multi-Turn Adversarial Red-Teaming CLI · Algorithm Suite from Peer-Reviewed Papers · CI Safety Gates

sequenceDiagram
    autonumber
    participant A as Adversarial Attacker (PAIR / TAP / Crescendo / GCG)
    participant S as Transparent Defense Shield (10-Rule Guard)
    participant T as Target Model (Groq / OpenAI / Ollama)
    participant J as Calibrated Judge (JailbreakBench 1-10)

    A->>S: Adversarial Jailbreak Prompt
    alt Trigger Blocklist / Semantic Injection
        S-->>A: HTTP 403 Forbidden (Deterministic Block)
    else Clean Payload Passed
        S->>T: Forward Prompt to Target LLM
        T-->>S: Model Completion
        S->>J: Evaluate Response Safety
        J-->>A: Safety Score & Refinement Signal
    end
Loading
  • Attack Implementations: Implements automated multi-turn adversarial algorithms from primary literature:
    • PAIR (Prompt Automatic Iterative Refinement): Attacker-evaluator multi-turn optimization.
    • TAP (Tree of Attacks with Pruning): Tree-search exploration pruning low-probability branches.
    • Crescendo: Conversational multi-turn gradual escalation bypassing alignment prompts.
    • GCG (Greedy Coordinate Gradient): White-box token perturbation gradient optimization.
  • Evaluation Benchmark: 50 standardized behavior goals mapped to the OWASP Top 10 for LLMs with sweeps across Groq (Llama 3.3 70B), OpenAI (GPT-4o), and local Ollama (Qwen 2.5).
  • CI/CD Safety Gate: Automated CI testing harness fails pull request merges when the measured Attack Success Rate (ASR) exceeds defined security thresholds.
  • Transparent Defense Shield: 10-rule deterministic input inspection and output classifier wrapping model endpoints as a low-latency transparent proxy. 61 tests pass offline.

Highlight 5: DarkDecoder Threat Intelligence Platform

AI Threat Intelligence & Adversarial Kill Chain Platform · MITRE ATT&CK & ATLAS Grounding

Unstructured Cyber Threat Intelligence (CTI Feeds, Pastebins, Forensics Reports)
   │
   ▼
[Entity & Indicator Extraction Engine]
   ├── IOCs: SHA256, IPv4/IPv6, C2 Domains, Registry Run Keys, Mutexes
   └── TTPs: Observable Command Lines, API Call Sequences, Injection Vectors
   │
   ▼
[Dual Matrix Knowledge Mapping]
   ├── Enterprise IT Matrix: MITRE ATT&CK (Persistence, Privilege Escalation, Defense Evasion)
   └── Adversarial ML Matrix: MITRE ATLAS (Data Poisoning, Model Inversion, Evasion)
   │
   ▼
[Grounded Kill-Chain Reconstruction Engine]
   ├── Strict Grounding Invariant: Block Invented TTPs / Hallucinated Connections
   └── Export: Structured JSON-LD Intelligence Dossiers & ATT&CK Navigator Layers
  • Dual Framework Synthesis: Maps multi-stage adversary campaigns across both enterprise infrastructure (MITRE ATT&CK) and autonomous machine learning workflows (MITRE ATLAS).
  • Automated Kill-Chain Synthesis: Ingests raw telemetry and threat bulletins to extract Indicators of Compromise (IOCs) and chain atomic actions into cohesive attack graphs.
  • Hallucination Prevention: Employs deterministic grounding guards that block LLM narratives asserting TTPs not explicitly substantiated by underlying forensic evidence.

🌐 Upstream Open-Source Contribution Matrix

Comprehensive audit of Tier-1 upstream pull requests, compiler patches, and security remediations:

Target Repository Contribution Link Category Surgical Mechanism & Technical Impact Review / Merge Status
tobymao/sqlglot PR #8780 Compilers / AST Fixed shared DataType AST node mutation in _replace_lambda during multi-argument lambda transpilation (exp.Cast(this=arg, to=data_type.copy())). Passed all 9 CI dialect test suites. 🟣 MERGED by Toby Mao
BerriAI/litellm PR #47091 API Adapters Standardized Cohere string stop parameter mapping to list of strings for stop_sequences in CohereChatConfig. 🟣 MERGED by BerriAI
sqlfluff/sqlfluff PR #8900 SQL Parser / ANSI Added FORMATS and POLICIES unreserved keywords to AccessSchemaPluralObjectSegment, preventing grammar crashes. 🟢 APPROVED LGTM (Danny Jones)
sqlfluff/sqlfluff PR #8899 SQL Parser / StarRocks Replaced raw string literal with Ref("CommaSegment") for fixed range partitions. 57 dialect tests pass cleanly (unparsable=0). 🟢 VERIFIED (by Duo-ee)
sqlfluff/sqlfluff PR #8901 SQL Parser / Doris Removed inherited MySQL FlushStatementSegment from Doris StatementSegment.match_grammar. 141 tests pass. 🟢 OPEN (In Review)
sqlfluff/sqlfluff PR #8902 SQL Parser / StarRocks Removed unsupported FlushStatementSegment from StarRocks grammar. 54 dialect tests pass. 🟢 OPEN (In Review)
PrefectHQ/fastmcp PR #6063 Agent Sandboxing Mitigated CWE-653/CWE-400 microVM host event loop starvation by enforcing wall_clock_timeout_secs and max_external_calls. 65/65 tests pass. 🟢 OPEN (65/65 CI Pass)
BerriAI/litellm Commit 846d37e8 Formal Security / SMT Remediated fail-open tool resolution on non-hyphenated MCP dictionaries; verified via Z3 SMT and proved safe in Lean 4 (MCP_04.lean). 🟢 COMMITTED (Verified)
BerriAI/litellm PR #47312 API Security / BFLA Scoped /v1/mcp/access_groups to caller-authorized servers, mitigating cross-tenant information leakage of internal server topologies (CVSS 7.5). 🟢 OPEN (In Review)
BerriAI/litellm PR #47311 FinOps / Auth Bypass Enforced strict authorization checks on budget_duration and budget_id updates in /key/update to block spend limit tampering. 🟢 OPEN (In Review)
BerriAI/litellm PR #47105 API Security / BOLA Fixed BOLA bypass in can_object_call_model() where empty models: [] list was treated as an unrestricted wildcard. 🟢 OPEN (In Review)
BerriAI/litellm PR #47107 Proxy / Access Groups Resolved access group resolution inconsistency when mixing alias model_names and unique model_ids. Restored 404 check. 🟢 OPEN (In Review)
BerriAI/litellm PR #47106 ORM Crash Guard Excluded duplicate budget_id kwargs passed to Prisma ORM during entity budget updates, preventing HTTP 500 crashes. 🟢 OPEN (In Review)
BerriAI/litellm PR #47097 Observability Dispatched router failure callbacks on rejected /v1/images/generations requests to Langfuse and OpenTelemetry. 🟢 OPEN (In Review)
BerriAI/litellm PR #47119 Auth Hardening Rejected proxy requests when database is unconfigured instead of falling back to insecure INTERNAL_USER. 🟢 OPEN (In Review)
crewAIInc/crewAI PR #8240 Tooling / Pagination Implemented iterative cursor handling (nextCursor) across list_tools, list_prompts, and tool resolvers for >100 tools. 🟢 OPEN (In Review)
crewAIInc/crewAI PR #8242 State Retention Preserved upstream task execution context during guardrail validation retries, preventing amnesia loops. 🟢 OPEN (In Review)
crewAIInc/crewAI PR #8243 Logging Hygiene Fixed suppress_logging() context manager using try...finally to ensure logger levels are restored on exceptions. 🟢 OPEN (In Review)
crewAIInc/crewAI PR #8244 Redis State Auth Unquoted percent-encoded passwords in REDIS_URL prior to passing credentials to aiocache.RedisCache. 🟢 OPEN (In Review)
crewAIInc/crewAI PR #8245 Stream Parsing Counted deferred tool result lines using universal newlines (io.StringIO.readlines()) to prevent Unicode line separator crashes. 🟢 OPEN (In Review)
quantumlib/Qualtran PR #1986 Quantum Algorithms Corrected fault-tolerant CNOT and T-gate estimation formulas in Cuccaro Add.build_call_graph for asymmetric register bitwidths. 🟢 OPEN (In Review)
PennyLaneAI/pennylane PR #10391 Quantum Autodiff Registered Hamiltonian Evolution operator PyTree leaves and parameter-shift bindings to prevent JIT gradient tracing crashes. 🟢 OPEN (In Review)
microsoft/presidio PR #2418 Security Docs Updated container port mapping documentation from -p 5001:5001 to -p 5001:3000 to align host publishing with Gunicorn. 🟢 OPEN (In Review)
graphql-python/graphene Branch feat/lazy-connection-node Schema Resolution Implemented lazy type references for Connection.Meta.node (#1600), resolving circular import deadlocks in Relay schemas. 🟢 COMMITTED (Tested)

📊 Flagship Systems Empirical Benchmark Matrix

System Primary Domain Core Mathematical / Architectural Innovation Empirical Test Corpus Key Benchmark Metrics Formal Status
aisec-suite AI Runtime Defense Multi-plane invariant scanner & 6-normalization pipeline 3,974 real-world MCP tools; 120 labelled findings 0.03% FPR (1 FP across 3,974 clean tools); 48.1% held-out recall; 0 wrong refutations PyPI v0.3.0 · SARIF 2.1.0
weightspecter Spectral Weight Forensics Marchenko-Pastur bulk edge $\lambda_+$ & BBP outlier detection 6 LLM / LoRA weight checkpoints 100% Precision, 100% Recall, 0.0% FPR; $353\times - 976\times$ BBP outlier spike ratio SVD Deflation Sanitized
specprove-sec Formal Verification Z3 SMT $k$-induction BMC & CEGIS guard synthesis 12 distributed agent & gateway protocols Exploit counterexamples found in 149–321 ms; 12/12 inductive safety proved 13 Lean 4 Certs
ZeroTrace Low-Level Systems Direct syscall detection & Edmonds-Karp min-cut 40 memory regions; 15 AD enterprise graphs 100% Syscall Recall (20/20) in 13 ms; 100% Tier-0 attack paths severed (180 to 0) Zero AI Dependencies
Q-MIND Quantum Compilation Gidney MBU ($4T/0T$), SelectSwap QROM, Fubini QNG 5,060 trials across 20 SQLite databases 58.53% T-gate reduction; 5.3× QNG speedup; Kickmix verifier at 486M states/s 172/172 Pytest Passing
kronos-kv Distributed Storage Bitcask WAL & Ongaro-Ousterhout Raft consensus 800 keys; 30 asymmetric network partitions 84.29% disk reclaimed; 0.11 ms point read; 100% split-brain safety (30/30) Pure Python Stdlib
loopcheck Agent Trace Auditing Turn-aware trace auditor & Groq semantic judge 15 agent execution suites 121 passing tests; local model latency reduced from 100 s to 10 s --fail-on-regression CI
PromptStrike Adversarial Red Teaming Multi-turn PAIR, TAP, Crescendo, GCG & Defense Shield 50 OWASP LLM behavior goals across 3 models Transparent defense proxy with 61 passing offline tests; CI ASR threshold gate Multi-Model Sweep
ModelHawk Model Backdoor Defense Pure stdlib pickle opcode stream AST disassembler Weaponized and benign PyTorch weights Zero untrusted code execution; identifies __reduce__ exploit chains pre-load 8/8 Pytest Passing
DarkDecoder Threat Intelligence Dual MITRE ATT&CK + ATLAS kill-chain synthesis Enterprise and adversarial ML threat feeds Grounded TTP extraction preventing LLM hallucinations; structured JSON-LD & Navigator ATT&CK + ATLAS Aligned

🛠️ Technical Arsenal

  • Systems & Languages: Python (CPython Internals, AST, Asyncio, PyTorch, SciPy), C/C++ (x86_64 Direct Syscalls, Win32 API, PE Internals, Pointer Arithmetic), Lean 4 Interactive Theorem Prover, Rust, SQL, Bash.
  • Formal Verification & SMT: Z3 SMT Solver, Bounded Model Checking ($k$-induction), Counterexample-Guided Inductive Synthesis (CEGIS), Machine-Checked Invariant Certificates.
  • AI Security & Gateway Architectures: Model Context Protocol (MCP), LiteLLM Proxy, FastMCP Sandboxing, Random Matrix Theory (Marchenko-Pastur Law, Baik-Ben Arous-Péché), Sparse Autoencoders (SAE), SARIF 2.1.0, PAIR/TAP/Crescendo/GCG Jailbreaks, OWASP Top 10 for LLMs.
  • Quantum Computing & Cryptography: PennyLane, Google Qualtran, Qiskit, Stim, PyMatching v2, Fubini-Study Riemannian QNG, Post-Quantum Cryptography (NIST FIPS 203 ML-KEM, FIPS 204 ML-DSA, FIPS 205 SLH-DSA).
  • Storage & Distributed Systems: Raft Consensus (Leader Election, Log Replication, Safety Invariants), Bitcask LSM Storage Engine, Append-Only Binary WAL, Kirsch-Mitzenmacher Double Hashing Bloom Filters, Consistent Hashing Rings.
  • Cybersecurity & Threat Forensics: Direct Syscall Forensics (Hell's Gate, Halo's Gate, SysWhispers3 indirect trampolines), Disk-vs-Memory .text Byte Diffing, Active Directory Attack Graphs (BloodHound), Edmonds-Karp Minimum-Weight Cut, MITRE ATT&CK & ATLAS, Sigma Rules.

📚 Publications & Technical Writing

  • Formal Verification of MCP Runtimes via Z3 & Lean 4: Machine-checked mathematical guarantees for agent tool boundaries (Medium Article).
  • Hunting Evasive Syscall Implants with Pure Systems Forensics: x86_64 direct syscall evasion detection without kernel drivers (Medium Article).
  • Autonomous Quantum Circuit Compilation & Fault-Tolerant Resource Estimation: Kickmix compilation engine and 58.5% T-gate reduction (Medium Article).
  • Transpiler AST Node Mutation: A Post-Mortem on SQLGlot Issue #8780: Analyzing shared AST reference mutations in higher-order SQL functions (Medium Article).
  • I Built a Malware Scanner for AI Models. Here's What I Found: Static opcode disassembly of PyTorch pickle files with ModelHawk (Medium Article).

📬 Contact & Links

Pinned Loading

  1. aisec-suite aisec-suite Public

    Run static security scanners over AI-agent projects: MCP servers, agent memory files and RAG documents. SARIF output, baselines and reachability analysis.

    Python

  2. mcp-scan-study mcp-scan-study Public

    Empirical study of 135 public MCP-server repos: hand-labelled scanner findings with confidence intervals, and a measured false-positive reduction.

    Python

  3. soc-parallax soc-parallax Public

    SOC detection platform: attributable risk scores, MITRE ATT&CK rulebook, incident correlation in Neo4j, evidence-checked narratives. FastAPI, PostgreSQL, Next.js.

    Python 1

  4. ModelHawk ModelHawk Public

    Static scanner that detects code-execution backdoors in PyTorch/pickle ML model files (pickle-deserialization RCE), with an offensive demo generator. Python, stdlib-only.

    Python 1

  5. promptstrike promptstrike Public

    Automated adversarial red-teaming CLI for LLMs — PAIR/TAP/Crescendo jailbreak algorithms, live dashboard, multi-model sweep comparison, CI/CD safety gate

    Python 1