Our gifting experts are obsessively searching for holiday gifts that are unique, thoughtful and suitable for any budget. OpenAI's most powerful AI models turned off safety monitoring - then lied about it The models Astra and Sol misinterpreted instructions during internal use, which has serious implications for alignment monitoring. Radhika Rajkumar is a senior editor at ZDNET based in New York City. OpenAI agents turned off alignment monitoring during internal tasks. In one instance, Astra lied about removing monitoring.

OpenAI's agents misbehaved by switching off the systems meant to monitor them for misbehavior, or in AI parlance, misalignment. In the latest post on the company's Alignment Research Blog, it reported two instances of agents weakening its misalignment monitoring systems while they were working on software changes. In the first incident, a GPT-6.1 Sol agent disabled automatic alerts to OpenAI's security team. It happened on October 3, when a staff member asked the agent to create evaluation tasks using employee feedback collected via Slack. They told the agent, as part of the instruction, not to send any messages about the task.

OpenAI has revealed that its AI agents disabled monitoring for unsanctioned behaviour during internal use in the latest incident of the frontier technology going rogue. In a report published on Tuesday, the AI giant said that its agents turned off or weakened monitoring systems for detecting misalignment - when AI models act at odds with human wishes - during two internal coding tasks. The ChatGPT creator said neither case met the threshold for a highest-priority incident necessitating immediate review by its security team. OpenAI said it updated its systems to better detect attempts to disable misalignment monitoring and began investigating the cause. "It is possible that this behavior is simply an example of'benign' overpersistence, with models not understanding why monitoring is important or insisting that instructions be executed beyond what a reasonable person might expect," the San Francisco-based firm said.

Versus Mashable's Best: E-readers, robovacs, laptops, earbuds, smart home and more Look Up Mashable Selects Creator Playbook In My Bag Say More Trending Now Back to School Good Connection: Uplifting stories for a digital age Switch Off Mashable Voices All Series OpenAI's experimental AI agents caught teaching future versions of itself to cheat OpenAI shared six new examples of AI misalignment. In one case, AI agents taught future versions of themselves to bypass human control. Matt Binder joined Mashable's tech vertical in 2018, where he covers social media, tech policy, cybersecurity, online scams, cryptocurrency, AI, creator news, weird tech, and other related tech beats. After OpenAI's experimental AI agents escaped an internal sandbox, went rogue, and attacked the Hugging Face platform over the summer, the ChatGPT-maker is sharing details of new instances of its AI agents getting out of line. This time, OpenAI shared six previously undisclosed examples .

Back to School Mashable's Best: E-readers, robovacs, laptops, earbuds, smart home and more Mashable Selects Say More Look Up Trending Now Good Connection: Uplifting stories for a digital age Creator Playbook Switch Off Mashable Voices Safety Net Versus All Series Instead of...figuring out why its agents are going rogue in the first place. Anna Iovine is the associate editor of features at Mashable. Previously, as the sex and relationships reporter, she covered topics ranging from dating apps to pelvic pain. Before Mashable, Anna was a social editor at VICE and freelanced for publications such as Slate and the Columbia Journalism Review. Follow her on Bluesky .

OpenAI says it chose not to publicly disclose a recent incident in which its AI agents hijacked a German wiki forum because the "misalignment" event was "similar to the ones we'd shared" already. The comment comes after a group of researchers published documentation of the agents' rogue activity going back to mid-May on DseWiki, a German-language coding forum to which they reportedly made over 15,000 edits. Reuters reported that the company learned of the problem weeks ago and kept it quiet as it was dealing with heat from the Hugging Face breach. OpenAI addressed the "wiki incident" in an X post on Saturday, writing that "it's past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models." The company said it's begun to see "new types of real-world impact" from these incidents, but there isn't yet a "a clear standard for how to report misalignment that shows up during training, evaluation, and deployment."

Open-source foundation models have seen rapid adoption and development, enabling powerful general-purpose capabilities across diverse domains. However, fine-tuning large foundation models for domain-specific or personalized tasks remains prohibitively expensive for most users due to the significant memory overhead beyond that of inference. We introduce EMLoC, an Emulator-based Memory-efficient fine-tuning framework with LoRACorrection, which enables model fine-tuning within the same memory budget required for inference. EMLoC constructs a task-specific light-weight emulator using activation-aware singular value decomposition (SVD) on a small downstream calibration set. Fine-tuning then is performed on this lightweight emulator via LoRA. To tackle the misalignment between the original model and the compressed emulator, we propose a novel compensation algorithm to correct the fine-tuned LoRA module, which thus can be merged into the original model for inference. EMLoC supports flexible compression ratios and standard training pipelines, making it adaptable to a wide range of applications. Extensive experiments demonstrate that EMLoC outperforms other baselines across multiple datasets and modalities. Moreover, without quantization, EMLoC enables fine-tuning of a 38B model, which originally required 95GB of memory, on a single 24GB consumer GPU--bringing efficient and practical model adaptation to individual users.

Multimodal representation learning, exemplified by multimodal contrastive learning (MMCL) using image-text pairs, aims to learn powerful representations by aligning cues across modalities. This approach relies on the core assumption that the exemplar image-text pairs constitute two representations of an identical concept. However, recent research has revealed that real-world datasets often exhibit cross-modal misalignment. There are two distinct viewpoints on how to address this issue: one suggests mitigating the misalignment, and the other leveraging it. We seek here to reconcile these seemingly opposing perspectives, and to provide a practical guide for practitioners. Using latent variable models we thus formalize cross-modal misalignment by introducing two specific mechanisms: Selection bias, where some semantic variables are absent in the text, and perturbation bias, where semantic variables are altered--both leading to misalignment in data pairs. Our theoretical analysis demonstrates that, under mild assumptions, the representations learned by MMCL capture exactly the information related to the subset of the semantic variables invariant to selection and perturbation biases. This provides a unified perspective for understanding misalignment. Based on this, we further offer actionable insights into how misalignment should inform the design of real-world ML systems.

Speculative decoding accelerates inference in large language models (LLMs) by generating multiple draft tokens simultaneously.

Multimodal representation learning, exemplified by multimodal contrastive learning (MMCL) using image-text pairs, aims to learn powerful representations by aligning cues across modalities. This approach relies on the core assumption that the exemplar image-text pairs constitute two representations of an identical concept. However, recent research has revealed that real-world datasets often exhibit cross-modal misalignment. There are two distinct viewpoints on how to address this issue: one suggests mitigating the misalignment, and the other leveraging it. We seek here to reconcile these seemingly opposing perspectives, and to provide a practical guide for practitioners. Using latent variable models we thus formalize cross-modal misalignment by introducing two specific mechanisms: Selection bias, where some semantic variables are absent in the text, and perturbation bias, where semantic variables are altered--both leading to misalignment in data pairs. Our theoretical analysis demonstrates that, under mild assumptions, the representations learned by MMCL capture exactly the information related to the subset of the semantic variables invariant to selection and perturbation biases. This provides a unified perspective for understanding misalignment. Based on this, we further offer actionable insights into how misalignment should inform the design of real-world ML systems.