Kellin Pelrine
Member of Technical Staff
FAR.AI
Kellin Pelrine leads FAR.AI's Integrity team, which aims to make AI both trustworthy and secure. The team works to uncover, understand, and solve AI risks across misuse, misalignment, and loss of control. Some of their current agendas span red-teaming, tampering, persuasion, evaluation, and demos. The Integrity team has discovered vulnerabilities and helped improve safeguards of models from every frontier AI company.
Previously, Kellin co-founded multiple startups, completed a PhD in AI at McGill University and the Mila AI Institute, and completed master’s degrees in Economics (Yale University) and Applied Mathematics (University of Colorado Boulder).
NEWs & publications
NEWs & publications
Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attribution
February 19, 2026
concept-data-attribution-02-2026
Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attribution
concept-influence-leveraging-interpretability-to-improve-performance-and-efficiency-in-training-data-attribution
Revisiting Frontier LLMs’ Attempts to Persuade on Extreme Topics: GPT and Claude Improved, Gemini Worsened
February 11, 2026
revisiting-attempts-to-persuade
Revisiting Frontier LLMs’ Attempts to Persuade on Extreme Topics: GPT and Claude Improved, Gemini Worsened
revisiting-attempts-to-persuade
Frontier LLMs Attempt to Persuade into Harmful Topics
August 21, 2025
attempt-to-persuade-eval
It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics
its-the-thought-that-counts-evaluating-the-attempts-of-frontier-llms-to-persuade-on-harmful-topics
Illusory Safety: Redteaming DeepSeek R1 and the Strongest Fine-Tunable Models of OpenAI, Anthropic, and Google
February 4, 2025
illusory-safety-redteaming-deepseek-r1-and-the-strongest-fine-tunable-models-of-openai-anthropic-and-google
Jailbreak-Tuning: Models Efficiently Learn Jailbreak Susceptibility
jailbreak-tuning-models-efficiently-learn-jailbreak-susceptibility
GPT-4o Guardrails Gone: Data Poisoning & Jailbreak-Tuning
October 31, 2024
gpt-4o-guardrails-gone-data-poisoning-jailbreak-tuning
Data Poisoning in LLMs: Jailbreak-Tuning and Scaling Laws
scaling-laws-for-data-poisoning-in-llms
Beyond the Board: Exploring AI Robustness Through Go
June 18, 2024
beyond-the-board-exploring-ai-robustness-through-go
Can Go AIs be adversarially robust?
can-go-ais-be-adversarially-robust
We Found Exploits in GPT-4’s Fine-tuning & Assistants APIs
December 21, 2023
we-found-exploits-in-gpt-4s-fine-tuning-assistants-apis
Exploiting Novel GPT-4 APIs
exploiting-novel-gpt-4-apis
Even Superhuman Go AIs Have Surprising Failure Modes
July 15, 2023
even-superhuman-go-ais-have-surprising-failure-modes
Adversarial Policies Beat Superhuman Go AIs
adversarial-policies-beat-superhuman-go-ais
TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering
February 6, 2026
tamperbench-systematically-stress-testing-llm-safety-under-fine-tuning-and-tampering
Prefill-level Jailbreak: A Black-Box Risk Analysis of Large Language Models
February 19, 2026
prefill-level-jailbreak-a-black-box-risk-analysis-of-large-language-models
Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attribution
February 19, 2026
concept-influence-leveraging-interpretability-to-improve-performance-and-efficiency-in-training-data-attribution
Concept Influence: Leveraging Interpretability to Improve Performance and Efficiency in Training Data Attribution
concept-data-attribution-02-2026
Revisiting Frontier LLMs’ Attempts to Persuade on Extreme Topics: GPT and Claude Improved, Gemini Worsened
February 11, 2026
revisiting-attempts-to-persuade
Revisiting Frontier LLMs’ Attempts to Persuade on Extreme Topics: GPT and Claude Improved, Gemini Worsened
revisiting-attempts-to-persuade
Large language models can effectively convince people to believe conspiracies
January 9, 2026
large-language-models-can-effectively-convince-people-to-believe-conspiracies
Emergent Persuasion: Will LLMs Persuade Without Being Prompted?
October 21, 2025
emergent-persuasion-will-llms-persuade-without-being-prompted
Open Technical Problems in Open-Weight AI Model Risk Management
October 1, 2025
open-technical-problems-in-open-weight-ai-model-risk-management
Jailbreak-Tuning: Models Efficiently Learn Jailbreak Susceptibility
July 15, 2025
jailbreak-tuning-models-efficiently-learn-jailbreak-susceptibility
Accidental Misalignment: Fine-Tuning Language Models Induces Unexpected Vulnerability
May 22, 2025
accidental-misalignment-fine-tuning-language-models-induces-unexpected-vulnerability
It's the Thought that Counts: Evaluating the Attempts of Frontier LLMs to Persuade on Harmful Topics
July 20, 2025
its-the-thought-that-counts-evaluating-the-attempts-of-frontier-llms-to-persuade-on-harmful-topics
Frontier LLMs Attempt to Persuade into Harmful Topics
attempt-to-persuade-eval
Illusory Safety: Redteaming DeepSeek R1 and the Strongest Fine-Tunable Models of OpenAI, Anthropic, and Google
February 4, 2025
illusory-safety-redteaming-deepseek-r1-and-the-strongest-fine-tunable-models-of-openai-anthropic-and-google
Illusory Safety: Redteaming DeepSeek R1 and the Strongest Fine-Tunable Models of OpenAI, Anthropic, and Google
illusory-safety-redteaming-deepseek-r1-and-the-strongest-fine-tunable-models-of-openai-anthropic-and-google
Data Poisoning in LLMs: Jailbreak-Tuning and Scaling Laws
August 6, 2024
scaling-laws-for-data-poisoning-in-llms
GPT-4o Guardrails Gone: Data Poisoning & Jailbreak-Tuning
gpt-4o-guardrails-gone-data-poisoning-jailbreak-tuning
Can Go AIs be adversarially robust?
June 18, 2024
can-go-ais-be-adversarially-robust
Even Superhuman Go AIs Have Surprising Failure Modes
even-superhuman-go-ais-have-surprising-failure-modes
Exploiting Novel GPT-4 APIs
December 21, 2023
exploiting-novel-gpt-4-apis
We Found Exploits in GPT-4’s Fine-tuning & Assistants APIs
we-found-exploits-in-gpt-4s-fine-tuning-assistants-apis
Adversarial Policies Beat Superhuman Go AIs
January 9, 2023
adversarial-policies-beat-superhuman-go-ais
Beyond the Board: Exploring AI Robustness Through Go
beyond-the-board-exploring-ai-robustness-through-go
