Ahsen Tahir

Final-year B.S. Computer Science student at FAST NUCES, graduating May 2027. I work on LLM post-training, Agentic AI, AI for software engineering, and AI safety.

Co-author on an accepted paper at the NeurIPS 2025 Workshop on Language Agents and World Models.

I am working on SASD (Source-Aware Self-Distillation), a training-time defense against prompt injection built on a multi-teacher KL divergence objective.

Mitacs Globalink Research Intern at the University of New Brunswick. I work in the AI + SE Research Lab with Prof. Francis Palma on the design-level security risks of coding-agent-generated microservice systems.

Applying to thesis-based MSc programs in Canada for Fall 2027, available to start September 2027. Reach me at ahsentahir007@gmail.com.

Research interests

AI for software engineering: security and quality of LLM-generated code, and evidence-grounded analysis of software systems
LLM post-training: instruction tuning, preference optimization, knowledge distillation
AI safety: keeping models faithful to user intent under untrusted input, prompt-injection defenses trained into the weights
AI security: building and shipping software securely in a world where models write it and act on it, including agent-generated systems
Agentic AI: coordination and communication among autonomous LLM agents
Evaluation: long-horizon agent tasks, and the security of code they generate

Ongoing work

SASD: Source-Aware Self-Distillation A training-time defense against prompt injection In progress · targeting NeurIPS / ICLR

Detectors and delimiters added at inference fall to an attacker who rewords the payload. SASD pushes source tracking into the weights: each teacher sees the task under a different source condition, and the student targets the one that never saw the injected span. Tested against adaptive attacks, not just fixed templates. Reproduced and benchmarked existing prompt-injection defenses on AgentDojo, establishing baselines across prior methods; the proposed approach already beats them. Work in progress. More.

Prompt injection
Knowledge distillation
AI safety

Detecting Design-Level Security Anti-Patterns in LLM-Generated Microservices Ahsen Tahir, Francis Palma, Shazil Khan Ongoing · University of New Brunswick

We study whether coding agents introduce security flaws at the architectural level of microservice systems. The project combines a literature-grounded catalog of nine design-level anti-patterns with an agentic detector that builds architecture diagrams and returns evidence-backed verdicts. More.

AI for software engineering
Software security
LLM agents

Publications

Anemoi: A Semi-Centralized Multi-agent System Based on Agent-to-Agent Communication MCP server from Coral Protocol NeurIPS 2025 Workshop on Language Agents and World Models (LAW) arXiv:2508.17068

Beyond Rule-Based Workflows: An Information-Flow-Orchestrated Multi-Agents Paradigm via Agent-to-Agent Communication from CORAL arXiv preprint, January 2026 arXiv:2601.09883

Full list, with author order and abstracts.

News

Jun 2026: Started as a Mitacs Globalink Research Intern at the University of New Brunswick, working with Prof. Francis Palma.
Feb 2026: Co-founded Velton, building an embedded copilot layer for B2B SaaS products.
Aug 2025: Anemoi accepted at the NeurIPS 2025 Workshop on Language Agents and World Models.
Feb 2025: Joined CoralOS as an AI Engineer, working on agent-to-agent communication infrastructure.