Geoff Salmon

Aaron Jaech, Adam Kalai, Adam Lerer et al. · openai

The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can reason about our safety policies in context when responding to potentially unsafe prompts, through deliberative alignment. This leads to state-of-the-art performance on certain benchmarks for risks such as generating illicit advice, choosing stereotyped responses, and succumbing to known jailbreaks. Training models to incorporate a chain of thought before answering has the potential to unlock substantial benefits, while also increasing potential risks that stem from heightened intelligence. Our results underscore the need for building robust alignment methods, extensively stress-testing their efficacy, and maintaining meticulous risk management protocols. This report outlines the safety work carried out for the OpenAI o1 and OpenAI o1-mini models, including safety evaluations, external red teaming, and Preparedness Framework evaluations.

Aaron Hurst, Adam Lerer, Adam P. Goucher et al. · openai

GPT-4o is an autoregressive omni model that accepts as input any combination of text, audio, image, and video, and generates any combination of text, audio, and image outputs. It's trained end-to-end across text, vision, and audio, meaning all inputs and outputs are processed by the same neural network. GPT-4o can respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds, which is similar to human response time in conversation. It matches GPT-4 Turbo performance on text in English and code, with significant improvement on text in non-English languages, while also being much faster and 50\% cheaper in the API. GPT-4o is especially better at vision and audio understanding compared to existing models. In line with our commitment to building AI safely and consistent with our voluntary commitments to the White House, we are sharing the GPT-4o System Card, which includes our Preparedness Framework evaluations. In this System Card, we provide a detailed look at GPT-4o's capabilities, limitations, and safety evaluations across multiple categories, focusing on speech-to-speech while also evaluating text and image capabilities, and measures we've implemented to ensure the model is safe and aligned. We also include third-party assessments on dangerous capabilities, as well as discussion of potential societal impacts of GPT-4o's text and vision capabilities.

Hazem M. Soliman, Geoff Salmon, Dušan Sovilj et al.

Advanced Persistent Threats (APTs) are sophisticated multi-step attacks, planned and executed by skilled adversaries targeting modern government and enterprise networks. Intrusion Detection Systems (IDSs) and User and Entity Behavior Analytics (UEBA) are commonly employed to aid a security analyst in the detection of APTs. The prolonged nature of APTs, combined with the granular focus of UEBA and IDS, results in overwhelming the analyst with an increasingly impractical number of alerts. Consequent to this abundance of data, and together with the crucial importance of the problem as well as the high cost of the skilled personnel involved, the problem of APT detection becomes a perfect candidate for automation through Artificial Intelligence (AI). In this paper, we provide, up to our knowledge, the first study and implementation of an end-to-end AI-assisted architecture for detecting APTs -- RANK. The goal of the system is not to replace the analyst, rather, it is to automate the complete pipeline from data sources to a final set of incidents for analyst review. The architecture is composed of four consecutive steps: 1) alert templating and merging, 2) alert graph construction, 3) alert graph partitioning into incidents, and 4) incident scoring and ordering. We evaluate our architecture against the 2000 DARPA Intrusion Detection dataset, as well as a read-world private dataset from a medium-scale enterprise. Extensive results are provided showing a three order of magnitude reduction in the amount of data to be reviewed by the analyst, innovative extraction of incidents and security-wise scoring of extracted incidents.

Hazem M. Soliman, Geoff Salmon, Dusan Sovilij et al.

Measuring similarity between IP addresses is an important task in the daily operations of any enterprise network. Applications that depend on an IP similarity measure include measuring correlation between security alerts, building baselines for behavioral modelling, debugging network failures and tracking persistent attacks. However, IPs do not have a natural similarity measure by definition. Deep Learning architectures are a promising solution here since they are able to learn numerical representations for IPs directly from data, allowing various distance measures to be applied on the calculated representations. Current works have utilized Natural Language Processing (NLP) techniques for learning IP embeddings. However, these approaches have no proper way to handle out-of-vocabulary (OOV) IPs not seen during training. In this paper, we propose a novel approach for IP embedding using an adapted graph neural network (GNN) architecture. This approach has the advantages of working on the raw data, scalability and, most importantly, induction, i.e. the ability to measure similarity between previously unseen IPs. Using data from an enterprise network, our approach is able to identify similarities between local DNS servers and root DNS servers even though some of these machines are never encountered during the training phase.