AI Jailbreak & Prompt-Injection Security Expert

What you'll do

You test AI models the way an attacker would, then turn what you find into something the safety team can act on. The ad names four areas: ethical jailbreaks, LLM red teaming, prompt injection and tool-use abuse. You design multi-turn, cross-domain approaches that expose bypass patterns a single prompt would miss.

The output is durable rather than one-off. You build and maintain regression test suites for jailbreak and injection susceptibility, construct evaluation frameworks that stress-test models against realistic adversarial threats, and write up methods and findings for both technical and non-technical readers. It is remote, contract-based work.

What you need

• 2+ years in adversarial machine learning, LLM red teaming, AI safety evaluation or a closely related security field
• Proven experience researching or uncovering ethical jailbreak, prompt-injection, tool-use abuse or adversarial AI vulnerabilities
• An advanced degree (PhD or MS) in computer science, cybersecurity or machine learning, or an equivalent professional background
• Clear writing and the ability to turn security findings into practical recommendations

Nice to have

• Recognition in the AI security community: published research, open-source tools or conference talks
• Familiarity with current LLM architectures, prompt engineering techniques and security assessment tools
• Experience on cross-functional AI safety projects

What it pays

micro1's listing states $50–90 per hour. The ad does not say whether pay is hourly or per task, how many hours a week are expected, or how a rate inside the band is set. Payment method and schedule are not published on the listing.

Worth knowing

Good:

• The stated experience bar is 2+ years, lower than the 5+ years micro1 asks for on its nuclear and military listings at the same band
• Regression suites and frameworks are lasting artefacts, so the work gives you something concrete to describe afterwards
• 20 openings, remote

Less good:

• Red-teaming can involve reading and producing unpleasant model output; the ad does not describe any support or content-exposure limits
• No hours, project length or start date are published, and the customer is not named
• Findings are written up for the project team, so public credit for the research is unlikely

About this listing

Posted by micro1 and read on 28 October 2026; the board showed 20 openings. We have not verified the pay band beyond what the ad states. If the role has closed since, the apply button will land on micro1's board rather than this posting. See micro1.