In recent years, Large Language Models (LLMs) have gained widespread use, raising concerns about their security. Traditional jailbreak attacks, which often rely on the model internal information or have limitations when exploring the unsafe behavior of the victim model, limiting their reducing their general applicability. In this paper, we introduce PathSeeker, a novel black-box jailbreak method, which is inspired by the game of rats escaping a maze. We think that each LLM has its unique "security maze", and attackers attempt to find the exit learning from the received feedback and their accumulated experience to compromise the target LLM's security defences. Our approach leverages multi-agent reinforcement learning, where smaller models collaborate to guide the main LLM in performing mutation operations to achieve the attack objectives. By progressively modifying inputs based on the model's feedback, our system induces richer, harmful responses. During our manual attempts to perform jailbreak attacks, we found that the vocabulary of the response of the target model gradually became richer and eventually produced harmful responses. Based on the observation, we also introduce a reward mechanism that exploits the expansion of vocabulary richness in LLM responses to weaken security constraints. Our method outperforms five state-of-the-art attack techniques when tested across 13 commercial and open-source LLMs, achieving high attack success rates, especially in strongly aligned commercial models like GPT-4o-mini, Claude-3.5, and GLM-4-air with strong safety alignment. This study aims to improve the understanding of LLM security vulnerabilities and we hope that this sturdy can contribute to the development of more robust defenses.

With the size and cost of large transformer-based language models growing, recently, there has been interest in shortcut casting of early transformer hidden-representations to final-representations for cheaper model inference. In particular, shortcutting pre-trained transformers with linear transformations over early layers has been shown to improve precision in early inference. However, for large language models, even this becomes computationally expensive. In this work, we propose Narrow Jump to Conclusions (NJTC) and Normalized Narrow Jump to Conclusions (N-NJTC) - parameter efficient alternatives to standard linear shortcutting that reduces shortcut parameter count by over 97%. We show that N-NJTC reliably outperforms Identity shortcuts at early stages and offers stable precision from all transformer block levels for GPT-2-XL, Phi3-Mini and Llama2-7B transformer models, demonstrating the viability of more parameter efficient short-cutting approaches.

The ripple effect poses a significant challenge in knowledge editing for large language models. Namely, when a single fact is edited, the model struggles to accurately update the related facts in a sequence, which is evaluated by multi-hop questions linked to a chain of related facts. Recent strategies have moved away from traditional parameter updates to more flexible, less computation-intensive methods, proven to be more effective in addressing the ripple effect. In-context learning (ICL) editing uses a simple demonstration `Imagine that + new fact` to guide LLMs, but struggles with complex multi-hop questions as the new fact alone fails to specify the chain of facts involved in such scenarios. Besides, memory-based editing maintains additional storage for all edits and related facts, requiring continuous updates to stay effective. As a result of these design limitations, the challenge remains, with the highest accuracy being only 33.8% on the MQuAKE-cf benchmarks for Vicuna-7B. To address this, we propose RippleCOT, a novel ICL editing approach integrating Chain-of-Thought (COT) reasoning. RippleCOT structures demonstrations as `newfact, question, thought, answer`, incorporating a thought component to identify and decompose the multi-hop logic within questions. This approach effectively guides the model through complex multi-hop questions with chains of related facts. Comprehensive experiments demonstrate that RippleCOT significantly outperforms the state-of-the-art on the ripple effect, achieving accuracy gains ranging from 7.8% to 87.1%.

Although Large Language Models (LLMs) have achieved remarkable performance in numerous downstream tasks, their ubiquity has raised two significant concerns. One is that LLMs can hallucinate by generating content that contradicts relevant contextual information; the other is that LLMs can inadvertently leak private information due to input regurgitation. Many prior works have extensively studied each concern independently, but none have investigated them simultaneously. Furthermore, auditing the influence of provided context during open-ended generation with a privacy emphasis is understudied. To this end, we comprehensively characterize the influence and hallucination of contextual information during summarization. We introduce a definition for context influence and Context-Influence Decoding (CID), and then we show that amplifying the context (by factoring out prior knowledge) and the context being out of distribution with respect to prior knowledge increases the context's influence on an LLM. Moreover, we show that context influence gives a lower bound of the private information leakage of CID. We corroborate our analytical findings with experimental evaluations that show improving the F1 ROGUE-L score on CNN-DM for LLaMA 3 by $\textbf{10}$% over regular decoding also leads to $\textbf{1.5x}$ more influence by the context. Moreover, we empirically evaluate how context influence and hallucination are affected by (1) model capacity, (2) context size, (3) the length of the current response, and (4) different token $n$-grams of the context. Our code can be accessed here: https://github.com/james-flemings/context_influence.

Now that OpenAI is becoming a for-profit company, it's making a tidy profit in the process. The Wall Street Journal reported that OpenAI has raised 6.6 billion in new funding from investors, nearly doubling its value to 157 billion. The new funding also makes it the largest venture capital deal in history. The new investors jumped on board after the artificial intelligence startup planned to switch from a charitable non-profit to a for-profit, product-focused company. If OpenAI fails to make the move to for-profit, investors have the right to pull their funding, according to Axios.

OpenAI has raised 6.6bn ( 5bn) in a funding round that values the artificial intelligence business at 157bn, with chipmaker Nvidia and Japanese group SoftBank among its investors. The San Francisco-based startup, responsible for the ChatGPT chatbot, did not give details of a reported restructuring that will transform it into a for-profit business. The funding round was led by Thrive Capital, a US venture capital fund, and other backers include MGX, an Abu Dhabi-backed investment firm. OpenAI's post-fundraising valuation puts it on a par with Uber, although it remains far below the 3tn level of its biggest backer of recent years, Microsoft, which also joined the fundraising. Other investors included Nvidia, a dominant player in the market for the chips that train and operate AI models, and Softbank, which counts the UK chip designer Arm among its investments.

It's little wonder that investors were clamoring to plow money into OpenAI. Alongside an announcement that the company had raised 6.6 billion in funding, OpenAI revealed that "every week, over 250 million people around the world use ChatGPT to enhance their work, creativity, and learning." That's a sharp rise since late August, when OpenAI said the chatbot had 200 million weekly users -- double the number it had last November. As of June, 350 million people were using OpenAI's tools each month, according to internal documents obtained by The New York Times. It's unclear how many people are paying for access versus those using the free tier.

Giving AI chatbots human feedback on their responses seems to make them better at giving convincing, but wrong, answers. The raw output of large language models (LLMs), which power chatbots like ChatGPT, often contains biased, harmful or irrelevant information, and their style of interaction can seem unnatural to humans. To get around this, developers often get people to evaluate a model's responses and then fine-tune it based on this feedback.

Some of the world's top AI labs suffer from inadequate safety measures--and the worst offender is Elon Musk's xAI, according to a new study. The French nonprofit SaferAI released its first ratings Wednesday evaluating the risk-management practices of top AI companies. Siméon Campos, the founder of SaferAI, says the purpose of the ratings is to develop a clear standard for how AI companies are handling risk as these nascent systems grow in power and usage. AI systems have already shown their ability to anonymously hack websites or help people develop bioweapons. Governments have been slow to put frameworks in place: a California bill to regulate the AI industry there was just vetoed by Governor Gavin Newsom.

We present BehAV, a novel approach for autonomous robot navigation in outdoor scenes guided by human instructions and leveraging Vision Language Models (VLMs). Our method interprets human commands using a Large Language Model (LLM) and categorizes the instructions into navigation and behavioral guidelines. Navigation guidelines consist of directional commands (e.g., "move forward until") and associated landmarks (e.g., "the building with blue windows"), while behavioral guidelines encompass regulatory actions (e.g., "stay on") and their corresponding objects (e.g., "pavements"). We use VLMs for their zero-shot scene understanding capabilities to estimate landmark locations from RGB images for robot navigation. Further, we introduce a novel scene representation that utilizes VLMs to ground behavioral rules into a behavioral cost map. This cost map encodes the presence of behavioral objects within the scene and assigns costs based on their regulatory actions. The behavioral cost map is integrated with a LiDAR-based occupancy map for navigation. To navigate outdoor scenes while adhering to the instructed behaviors, we present an unconstrained Model Predictive Control (MPC)-based planner that prioritizes both reaching landmarks and following behavioral guidelines. We evaluate the performance of BehAV on a quadruped robot across diverse real-world scenarios, demonstrating a 22.49% improvement in alignment with human-teleoperated actions, as measured by Frechet distance, and achieving a 40% higher navigation success rate compared to state-of-the-art methods.