Research.

[ AI-RED-TEAM ]

Adversarial evaluation

Probing alignment and safety failures in production LLMs and multi-agent systems.

• › Black-box and grey-box red-teaming of frontier LLM deployments
• › Multi-turn jailbreak analysis and failure-mode taxonomy development
• › Refusal consistency and policy-drift measurement across model versions
• › Adversarial prompts against safety classifiers, content moderation, and alignment training pipelines

[ MODEL-SEC ]

Model security & integrity

Abliteration, fine-tuning attacks, poisoning, and supply-chain risk.

• › Abliteration and refusal-removal technique analysis on open-weight models
• › Fine-tuning attacks — how alignment degrades under adversarial training data
• › Training-data poisoning and backdoor research on locally-trained models
• › Supply-chain risk across HuggingFace, GGUF, Ollama, and other model distribution channels

[ AGENT-SEC ]

Agentic system hardening

The new attack surface: tool-use, MCP, autonomous workflows, prompt injection.

• › Prompt injection in production agent stacks (tool-use, RAG, web-browsing agents)
• › MCP server security — auth boundaries, tool poisoning, capability creep
• › Autonomous workflow auditing — what can the agent reach? what can it exfiltrate?
• › Confused-deputy and cross-context contamination in multi-agent systems

[ OFFENSIVE-R&D ]

Offensive security R&D

Traditional pentest research, payload engineering, lab work on owned hardware.

• › Custom payload development and detection-evasion research
• › Network and protocol-level offensive tooling
• › AI-augmented offensive workflows — where models meaningfully accelerate red-team work
• › Dedicated lab environment: workstation-class compute, GPU inference rig, isolated network segment

// collaborate

Working on something adjacent?

Research collaborations, responsible-disclosure coordination, and paper review — reach out at dev@aesirkode.com.