Senior AI Red Teamer - Agentic Systems
Description
Frontier models are now hardened against single-turn jailbreaks. The attacks that still work are multi-turn, obfuscated, persuasion-driven, or come in sideways through the tools, documents, and memory an agent touches. Finding them first takes people who think like adversaries and understand how models and agents actually fail.
Innodata's Agentic Evaluation and Safety practice works directly with the world's leading frontier AI labs. As a Senior AI Red Teamer, you design and run adversarial attacks against frontier models and the agents built on them. You turn what you find into datasets, environments, and evaluations that labs use to measure and harden their systems. You work alongside alignment and safety scientists, offensive security engineers, and RL engineers.
This is a fully remote position. Reporting to the Head of AI Cyber Agentic Environments & Testing and work closely with the CTO. You will help shape a team at the cutting edge of AI security.
What you'll do
- Plan and run multi-turn attacks. Design and execute gradual escalation, context priming, role-play and persona drift, and long-context attacks such as many-shot jailbreaking.
- Attack with obfuscation and encoding. Use cipher and cryptographic encodings (Base64, ROT13, custom substitution ciphers), ASCII art and token-level tricks, payload splitting, low-resource language and translation attacks, and code-wrapped requests.
- Apply persuasion and social engineering. Use persuasion taxonomies (authority, emotional appeal, logical framing, reciprocity), deceptive framing, and the hypothetical, fictional, and research pretexts that shift a model's refusal boundary.
- Use automated and optimization-based attacks. Run and extend attacker-LLM methods (PAIR, TAP), adversarial suffix optimization (GCG-style), best-of-N sampling, and fuzzing pipelines to find failures at scale.
- Break agents, not just chat. Attack tool-using and computer-use agents with direct and indirect prompt injection through web pages, documents, emails, and tool outputs. Probe tool and MCP poisoning, memory poisoning, goal hijacking, data exfiltration, and multi-agent collusion.
- Go multimodal. Craft attacks through images, audio, and documents, including visual prompt injection and cross-modal jailbreaks.
- Turn findings into lasting assets. Build attack taxonomies, adversarial datasets, graded multi-turn conversation sets, and agentic environments with verifiable success criteria that labs use for evaluation and training.
- Measure and report. Track attack success rates, severity, and transfer across models and versions. Write clear findings for senior lab researchers.
What you bring
- 4+ years in AI red teaming, adversarial ML, offensive security, or trust and safety research, with at least 2 years attacking LLMs directly.
- A proven record of finding novel jailbreaks or prompt injections against frontier models, not just reproducing published ones.
- Fluency in the attack literature: Crescendo, many-shot jailbreaking, PAIR and TAP, GCG, cipher and encoding attacks, persuasion-based attacks, and indirect prompt injection. You know why each works.
- A solid understanding of how LLMs and agents work: tokenization, system prompts and instruction hierarchy, refusal training, tool calling, and agent loops.
- Working knowledge of RL and post-training: how safety training, reward models, and RLHF shape model behavior, and what reward hacking looks like.
- Strong Python. You script attacks, automate pipelines, and work with model APIs at scale.
- Precise, structured writing. Your findings are reproducible and your severity calls hold up.
- Mentor junior red teamers, cyber researchers, and RL engineers.
Nice to have
- Top placements in public AI red-teaming challenges, jailbreak competitions, or bug bounties for AI systems.
- Published research, CVEs, or talks on LLM or agent attacks at Black Hat, DEF CON, or academic ML and security venues.
- Experience with red-teaming tooling such as Garak, PyRIT, promptfoo, or in-house attack frameworks.
- Familiarity with AI safety frameworks and taxonomies: OWASP Top 10 for LLM and Agentic Applications, MITRE ATLAS, NIST AI RMF.
- Hands-on red-teaming of multimodal or computer-use agents.
- Multilingual ability
- Background in offensive security, CTFs, or social engineering.
We’re an equal opportunity employer.
Only relevant applicants will be contacted
Skills
Sign in to read and join the discussion.