Making AI systems break in the lab not in production
adversarial testing, red-teaming & evaluation.
I'm Khalil Jaouani, an AI security researcher based in Tunisia. I focus on probing language models and agent systems for the failure modes that matter jailbreaks, tool-use exploits, and quiet policy bypasses , then turn what I find into repeatable defenses.
LET'S TALKI'm a security researcher focused on making AI systems safe to actually ship not just safe on paper.
My work spans adversarial ML, LLM red-teaming, agent security, and evaluation tooling. Across every engagement, I'm drawn to the same question:
How do you turn a one-off exploit into a test the whole team can trust?
I enjoy the adversarial mindset , thinking like an attacker, then building the harness that keeps that attack from ever shipping again.
Four disciplines, one goal: models and agents that hold up when someone is actually trying to break them.
GET STARTED ↗Studying how models fail under crafted input — perturbations, prompt injection, distribution shift.
Structured attack campaigns against language models and agentic systems to surface jailbreaks and exploits.
Automated harnesses that turn one-off findings into regression tests run on every model update.
Coordinating with vendors and labs on remediation timelines and write-ups that inform without enabling misuse.
Certifications, current studies, and hands-on security work shaping how I approach infrastructure, detection engineering, and AI security.
ITeam University
Currently studying AI Engineering, deepening my foundations in machine learning, software engineering, and secure AI systems.
Currently studyingSecurity Engineer · SAIPH
December 2025 — Present
Security Engineer · Felbled Platform
April 2025 — December 2025
Security engineering, SOC operations, vulnerability remediation, SIEM, IDS, and incident triage.
Building detection, response, and application security capabilities across production environments.
Built and productionized a Dockerized SOC email-analysis platform automating suspicious email parsing, vector embeddings, MITRE ATT&CK mapping, and safe sandbox execution. Perform security engineering across infrastructure and application layers, including vulnerability identification, remediation planning, and security posture improvement.
Built and managed a full SOC environment for continuous security monitoring, configured firewalls, conducted infrastructure vulnerability scans, and led remediation efforts. Deployed Wazuh SIEM and Suricata IDS, monitored alerts, and performed incident triage.
A few pieces of work that show how I approach red-teaming, evaluation, and disclosure.
Found that mem0's identity-scoping filters (user_id/agent_id) only checked for the presence of a key, not the safety of its value — a wildcard filter bypassed tenant isolation entirely, and an unvalidated delete_all("*") call could wipe every tenant's memory in a shared collection with one call. Reported via GitHub Security Advisory GHSA-9f59-36vq-gcw6, accepted and credited by the maintainers.
A high-performance Go tool combining passive Certificate Transparency reconnaissance with active DNS brute-forcing, multi-record lookups, wildcard detection, false-positive filtering, and adjustable concurrency up to 200 threads.
View on GitHub ↗A web-based CyberChef alternative for encoding, decoding, hashing, and cryptographic operations, built with vanilla JavaScript, drag-and-drop file input, and zero backend dependencies.
View on GitHub ↗Have a model, agent, or pipeline you'd like stress-tested — or a disclosure you need coordinated? I'd like to hear about it.