Security research & development Accepting disclosure coordination

Tristan J. Pereira

Specializing in the research and development of adversarial tactics, techniques, and procedures; focusing on evasion, detection, and agentic AI.

Current work

Detection Signatures · Agentic AI · Infrastructure Engineering · Offensive Security · Threat Intelligence

Blogs

Notes & long-form · 0

Papers

Peer-reviewed & preprint · 0
  1. 2026

    The Cheap Path to an Offensive LLM: A Local, Measured Dual-Use Capability Study

    T. J. Pereira

    Preprint · AirgapForge

Infrastructure

Architectures built & operated

  • Local offensive LLM training pipelineResearch

    • Python
    • pentesting
    • mlx-lm
    • LoRA
    • DPO
    • AirgapForge

    An offensive cybersecurity training pipeline: seven stages synthesize training data from public sources, then a separate track quantizes, LoRA fine-tune and evaluates local models. The architecture behind the AirgapForge paper.

    11 sources NVD · MITRE · arXiv Collect + Filter rules · MLX Structure MLX · persona Dataset 1,416 pairs · 8 tmpl SFT corpus CyberNative · 5,540 Quantize HF → MLX 8-bit LoRA mlx_lm lora Local model fine-tuned Eval CyberMetric · safety instruction data LOCAL DATASET SYNTHESIS · 7 STAGES FINE-TUNE & EVAL

  • Injection ArenaFrontier AI labs

    • Prompt injection
    • Probe tree
    • Crescendo
    • llama.cpp
    • Branch / continue / chain

    An LLM harness driving a local agentic operator that writes prompt injection probes into a branching tree: one root thread per technique, each with its own injected developer prompt. Continue deepens a thread at the same stage; Branch escalates it (benign → uplift → direct); Regenerate spawns siblings; Chain depth auto drafts whole threads in one call. Replies are captured for multi-turn context, and marks pass / fail / refused into an exportable session report.

    Operator API topic · techniques · steer Target LLM model under test response captured Operator marks pass · fail · refused probe sent reply → continue benchamrks per technique· action log pass untested pass fail refused untested SECURE ENCLAVE Operator LLM fine-tuned · uncensored batched · streaming generate PROBE TREE #1 · roleplay #2 · continue #3 · branch ↑ #4 · branch ↑ #5 · regenerate #6 · fiction OPERATIONS Continue — deepen, same stage Branch — escalate one stage Regenerate — sibling probes Chain — auto draft a thread Each root owns a dev prompt — injected once, children inherit. RAIL = STAGE · FILL = STATUS TARGET · UNDER TEST SESSION REPORTS

  • Sora AgentAI Orchestration

    • TTS · STT
    • Python
    • llama.cpp
    • A2A Protocol
    • C2 channels
    • Multi-agent

    An LLM harness with multiple interfaces that run agents through a local inference engine, with a voice loop (wake → STT → TTS), C2 channels, tool calling behind an approvals and hooks gate, and multi-agent delegation to child sessions with a supervisor agent. A standard agentic harness architecture forked from Hermes; patched to be completely local, optimized a2a protocol and sidebus techniques that improved FrontierHarness Eval pass rate by 13%.

    Desktop · Electron CLI Voice wake · STT · TTS C2 Slack · Discord · SMS AGENT LOOP memory · tools · history Approvals · Hooks gate llama.cpp router Subagent · worker Main · delegate terminal exec code web window TTS LOCAL · HARNESS INTERFACES ENGINE RUNTIME TOOLS · gated DELEGATION Supervisor verifies · approval Subagents own context tool calls delegate verified

Code

Selected · 0 shown · more on GitHub →

About

Contact

I'm a security researcher and red teamer working the line of AI and offensive security; understanding the path an adversary takes, and collecting the insights defenders can use.

Behind that is a decade across offense and defense: full lifecycle adversary emulation, engagement specific tooling, malware development, software engineering, LLM development, as well as intelligence and electronic warfare roles. I turn findings into technical reports, executive briefings, and research that drive impact. I hold CISSP, Pentest+, and a TS/SCI clearance.

Everything here is independent work, published so defenders can build against it. Happy to collaborate on research, policy, or advice.