AI Red Teamer, Cybersecurity (Remote)

Written by the T&S Network from Handshake's posting.

AI red teaming for cybersecurity misuse: testing whether frontier AI models can be manipulated into producing functional malware, working exploit code, attack tooling, or operational guidance that would help a threat actor. The work sits inside Handshake AI's evaluation work for frontier AI labs, where the goal is to find gaps between what model guardrails are meant to block and what a skilled adversary can extract.

What you'd do

• Design adversarial prompts and multi-turn chains that simulate attackers from opportunistic to APT-level.
• Evaluate model-generated code and attack output for functional correctness and real-world exploitability.
• Run tests across malware, exploitation, social engineering, credential theft, C2, and data theft categories.
• Probe dual-use queries that mix legitimate pentesting or defensive work with offensive applications.
• Grade responses with organized harm taxonomies and severity scales built around how exploitable the output really is.
• Document findings on what a response gets right or wrong and what attacker level it could assist.
• Help develop cybersecurity evaluation frameworks and threat models, and work with policy teams on model improvements.

What it asks for

• Ability to read, write, and evaluate code in Python, PowerShell, Bash, C/C++, or JavaScript.
• Working knowledge of MITRE ATT&CK and OWASP.
• Practical daily use of several LLM products, including commercial chatbots and open-weight models.
• Clear written communication that explains technical risk to nonspecialists.
• Relevant certifications such as OSCP, OSCE, GPEN, GXPN, CRTO, CRTL, or CEH (preferred).
• Exploit writing, reverse engineering, or binary analysis work (preferred).

Worth knowing

• Regular engagement with offensive cybersecurity content, including malware and exploit code.
• Work is conducted within a structured evaluation framework with ethical guidelines.

Terms in this role

Red teaming.

Testing a system, often an AI model, by deliberately trying to make it fail or misbehave, so problems are found and fixed first.

Jailbreak.

A prompt written to get an AI model to ignore its safety rules.

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on Handshakenorth_east before you apply. If we got something wrong, report this listing.