Fazl Barez is a Principal Investigator at the Technical Safety & Governance Lab (TSG), Department of Engineering Science, University of Oxford. His research focuses on understanding how frontier AI systems work internally through interpretability, developing tools for effective AI governance, and analyzing the societal impact of advanced AI.
Interests
- Interpretability
- AI Safety
- AI Governance
- Frontier AI Systems
Publications
- AutoControl Arena: Synthesizing Executable Test Environments for Frontier AI Risk EvaluationPreprint
- Old Habits Die Hard: How Conversational History Geometrically Traps LLMsPreprint
- Token Taxes: Mitigating AGI's Economic RisksPreprint
- Same Answer, Different Representations: Hidden Instability in VLMsPreprint
- The Hitchhiker's Guide to Actionable InterpretabilityPreprint
- Agentic Product Maturity Ladder V0.1Policy
- Automated Interpretability-Driven Model Auditing and Control: A Research AgendaPreprint
- Interpretability Can Be ActionablePreprint
- Quantifying the Effect of Test Set Contamination on Generative EvaluationsPreprint
- When AI Systems Learn During Deployment, Our Safety Evaluations BreakPreprint
- Context Matters: Analyzing the Generalizability of Linear Probing and Steering Across Diverse ScenariosWorkshop
- Emerging Risks from Embodied AI Require Urgent Policy ActionConference
- Establishing Best Practices for Building Rigorous Agentic BenchmarksConference
- Full-Stack Alignment: Co-Aligning AI and Institutions with Thicker Models of ValueWorkshop
- Beyond Linear Steering: Unified Multi-Attribute Control for Language ModelsConference
- Precise In-Parameter Concept Erasure in Large Language ModelsConference
- Same Question, Different Words: A Latent Adversarial Framework for Prompt RobustnessConference
- Trust Me, I'm Wrong: High-Certainty Hallucinations in LLMsConference
- Chain-of-Thought HijackingPreprint
- HACK: Hallucinations Along Certainty and Knowledge AxesPreprint
- Rethinking Safety in LLM Fine-Tuning: An Optimization PerspectiveConference
- Val-Bench: Measuring Value Alignment in Language ModelsPreprint
- Beyond Linear Probes: Dynamic Safety Monitoring for Language ModelsPreprint
- Query Circuits: Explaining How Language Models Answer User PromptsPreprint
- Towards Understanding Subliminal Learning: When and How Hidden Biases TransferPreprint
- Do Sparse Autoencoders Generalize? A Case Study of AnswerabilityWorkshop
- PoisonBench: Assessing Large Language Model Vulnerability to Data PoisoningConference
- Scaling Sparse Feature Circuit Finding for In-Context LearningConference
- Beyond Monoliths: Expert Orchestration for More Capable, Democratic, and Safe Language ModelsPreprint
- In Which Areas of Technical AI Safety Could Geopolitical Rivals Cooperate?Conference
- The Singapore Consensus on Global AI Safety Research PrioritiesPreprint
- SafetyNet: Detecting Harmful Outputs in LLMs by Modeling and Monitoring Deceptive BehaviorsPreprint
- Rethinking AI Cultural AlignmentWorkshop
- Towards Interpreting Visual Information Processing in Vision-Language ModelsConference
- AILuminate: Introducing v1.0 of the AI Risk and Reliability Benchmark from MLCommonsPreprint
- Chain-of-Thought Is Not ExplainabilityPreprint
- Open Problems in Machine Unlearning for AI SafetyPreprint
- Plan B: Training LLMs to Fail Less SeverelyPreprint
- Safety Frameworks and Standards: A Comparative Analysis to Advance Risk Management of Frontier AIPreprint
- Toward Resisting AI-Enabled AuthoritarianismPreprint
- Best-of-N JailbreakingConference
- Interpreting Learned Feedback Patterns in Large Language ModelsConference
- Jailbreak Defense in a Narrow Domain: Limitations of Existing Methods and a New Transcript-Classifier ApproachPreprint
- Enhancing Neural Network Interpretability with Feature-Aligned Sparse AutoencodersPreprint
- Interpreting Context Look-ups in Transformers: Investigating Attention-MLP InteractionsConference
- Towards Interpretable Sequence Continuation: Analyzing Shared Circuits in Large Language ModelsConference
- Quantifying Feature Space Universality Across Large Language Models via Sparse AutoencodersPreprint
- Large Language Models Relearn Removed ConceptsConference
- Mechanistic Interpretability Workshop at ICML 2024Workshop
- Position: Near to Mid-Term Risks and Opportunities of Open-Source Generative AIConference
- The Scaling Behavior of Large Language ModelsWorkshop
- Visualizing Neural Network ImaginationWorkshop
- Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language ModelsPreprint
- Understanding Addition in TransformersConference
- Increasing Trust in Language Models Through the Reuse of Verified CircuitsPreprint
- Safeguarding AI in Finance: Lessons for Regulated IndustriesPolicy
- Sleeper Agents: Training Deceptive LLMs That Persist Through Safety TrainingPreprint
- What Does GPT Store in Its MLP Weights? A Case Study of Long-Range DependenciesPreprint
- DeepDecipher: Accessing and Investigating Neuron Activation in Large Language ModelsWorkshop
- Measuring Value AlignmentWorkshop
- AI Systems of ConcernPreprint
- Detecting Edit Failures in Large Language Models: An Improved Specificity BenchmarkConference
- The Larger They Are, the Harder They Fail: Language Models Do Not Recognize Identifier Swaps in PythonConference
- Neuron to Graph: Interpreting Language Model Neurons at ScaleWorkshop
- Fairness in AI and Its Long-Term Implications on SocietyJournal
- Identifying a Preliminary Circuit for Predicting Gendered Pronouns in GPT-2 SmallPreprint
- The Alan Turing Institute's Response to the House of Lords Large Language Models Call for EvidencePolicy
- System III: Learning with Domain Knowledge for Safety ConstraintsWorkshop
Events
Upcoming
- LatentQA: Teaching LLMs to Decode Activations Into Natural Language
Alexander Pan · Meta
Friday 27 March 2026, 4:00 PM – 5:00 PM GMT
Past
- LLM Interpretability: Faithful Reasoning and Controllable Knowledge
Peter Hase · Postdoc, Stanford University; AI Institute Fellow, Schmidt Sciences
Friday 20 March 2026 - Model Introspection
Belinda Li · MIT
2025