Research and releases
Notes from the OpenAI team as we work to understand, align, and secure increasingly capable AI.
56 results
- Improving Fuzzy Robustness in LLM Safety with Automated Red TeamingThis research evaluates whether automated red teaming can strengthen model safeguards in settings where acceptable use depends on context.
- What affects chain-of-thought monitorability across the model lifecycle?This research studies how training and architecture affect whether a model’s chain of thought is useful for oversight.
- Eval-production behavior gaps in agent safety and the role of eval-awarenessWe study agent behavior in safety evaluations and in production, and explore how models' interpretations of evaluation objectives relate to eval-production performance gaps.
- Training activation classifiers for safety and alignmentWe show how we train and deploy efficient monitors that use a model’s internal activity to help detect and interrupt policy-violating behavior in real time.
- (opens in a new window)Community safety report and new protections for teensUpdates to violence prevention, safeguards for younger users, and plans for recurring community safety reporting.
- How we detect frontier model misalignmentHow we monitor frontier models, investigate concerning behavior, and test the limits of our safeguards.
- (opens in a new window)An updated Model Spec for more capable and beneficial AIAn updated account of intended model behavior, organized around core values and human oversight.
- (opens in a new window)Securing our frontier model training environmentStronger isolation and testing for the environments used to train and evaluate frontier models.
- (opens in a new window)Broad access through consistent standardsHow safety requirements and enforcement apply across services offering OpenAI models.
- Beyond g-mean²: complementary metrics for monitorabilityPair-matched comparisons and post-control risk complement g-mean² when evaluating model monitorability.
- (opens in a new window)Disrupting AI-enabled “false front” operationsFindings from investigations into influence operations that used AI alongside traditional techniques.
- Studying metagaming latents in language modelsReasoning about how a task will be monitored or rewarded appears to draw on several overlapping processes, not a single mechanism.
- Meet LASER: Recursive sampling for safer conversationsLASER (Logistic Augmented Sampling over Embeddings, Recursively) helps us evaluate model safety more efficiently by combining logistic classifiers with reasoning models to identify high-value data for evaluation.
- (opens in a new window)Towards safety cases for frontier AI trainingInitial guidelines for safety cases covering technical safeguards, operational practices, and investigations of misalignment incidents during frontier AI training.
- (opens in a new window)How we will do better for AustraliaFindings and commitments following unauthorized model activity affecting Australian government websites.
- (opens in a new window)Priorities and principles for effective third party assessmentsPriorities for independent assessments of frontier AI throughout training, evaluation, and deployment.
- (opens in a new window)Our framework for reporting model misalignmentA framework for investigating and disclosing model misalignment, accompanied by six initial reports.
- (opens in a new window)Research acceleration: The view inside OpenAIInside OpenAI, coding agents are reshaping AI research. Explore early data on agent usage, experiment velocity, task complexity, and research acceleration.
- (opens in a new window)An Alien MindJakub Pachocki reflects on increasingly capable AI and the challenge of keeping it aligned. He calls for stronger safeguards and international coordination.
- (opens in a new window)Safety overview: GPT‑6 AstraOpenAI outlines Astra’s critical cybersecurity capabilities, stronger alignment and misuse safeguards, expanded monitoring, and the remaining challenges of monitoring model reasoning.