AI News Archive: July 21, 2026 — Part 12
Sourced from 500+ daily AI sources, scored by relevance.
- Gradient-guided adapter merging for neuroimaging vision-language models
Automated interpretation of neuroimaging studies requires simultaneous assessment of multiple imaging evidence variables, each tied to distinct anatomical structures. Vision-language models (VLMs) offer a unified framework for multi-task analysis, but adapting pre-trained VLMs remains challenging. Full fine-tuning is computationally prohibitive, and joint multi-task training requires simultaneous access to all task data, which is often infeasible in clinical settings. Although model merging enables multi-task composition without joint re-training, existing methods focus on post-hoc algorithms with limited extension to VLMs and minimal application to neuroimaging. Here, we present GRadient-guided Adapter Merging (GRAM), a layer-selective low-rank adaptation (LoRA)-based fine-tuning and merging framework for multi-task neuroimaging visual question-answering (VQA). GRAM uses a gradient ratio that contrasts class-specific gradients to identify task-discriminative layers, and applies subspace-constrained projected gradient descent to restrict LoRA updates to directions consistent with the geometry of the pre-trained model. We leveraged a structured VQA benchmark, developed from the National Alzheimer's Coordinating Center (NACC) dataset, that pairs multi-sequence brain MRI studies with question-answer pairs across clinically relevant imaging evidence variables. Experiments on the VQA benchmark showed that GRAM outperformed or matched all-layer LoRA fine-tuning and a standard merging baseline while reducing inter-task interference during merging, and approached or surpassed the performance of joint multi-task training without joint re-training.
- Quantitative Fundus Autofluorescence in Early Dry AMD Using ImageJ: Near-Perfect Interobserver Agreement and Pattern-Specific Intensity Characterization
Purpose: To assess the feasibility of quantitative fundus autofluorescence (FAF) measurement in early age-related macular degeneration (AMD) using the freely available ImageJ software, to characterize signal intensity across FAF patterns, and to evaluate interobserver reproducibility in pattern classification. Methods: Single-center, non-blinded, retrospective, consecutive-case analytical study. FAF images acquired with Spectralis OCT+HRA (Heidelberg Engineering) from patients with early dry AMD seen at a tertiary referral center between January 2010 and September 2016 were analyzed. A standardized 300x300-pixel region of interest (ROI) centered on the fovea was evaluated in ImageJ v2.0.0-rc54/1.51h (Fiji distribution). Mean, minimum, and maximum autofluorescence (AF) pixel intensity were recorded. Each image was independently classified according to the Bindewald classification system by two graders; a third senior grader adjudicated discordances. Cohen's kappa (k) was used to assess interobserver agreement. Results: Of 423 patients with available FAF studies, 107 had dry AMD; 45 met quality and diagnostic criteria for early AMD and were included in the quantitative analysis. Mean age was 73.47 +/- 8.1 years; 62.2% were female. Mean FAF intensity was 120.26 (range 74.76-160.79); mean minimum was 32.07 (range 3-63) and mean maximum was 205.80 (range 125-255). Seven of eight Bindewald patterns were identified; the stippled pattern was absent. The most frequent pattern was minimal changes (31.1%), followed by increased focal (24.4%) and patchy (15.6%). Reticular pattern showed the highest mean AF (143.8), while lace pattern showed the lowest (88.4). Interobserver agreement for Bindewald pattern classification was almost perfect (k = 0.969; 95% CI, 0.908-1.000; p < 0.001). Agreement for lesion extent was moderate (k = 0.531) and for foveal involvement was substantial (k = 0.622). Conclusions: Quantitative FAF evaluation of early AMD using ImageJ is feasible and reproducible. ImageJ represents a cost-free alternative for multimodal retinal image analysis, with potential for automated screening applications in resource-limited settings. Keywords: age-related macular degeneration; fundus autofluorescence; ImageJ; quantitative autofluorescence; image analysis; Bindewald classification; interobserver agreement
- AI Couple Photo Maker | CreateVision
Merge two selfies into one realistic romantic couple photo, free.
- Multi-model forecasting of respiratory disease activity in Germany during the 2024-2025 season
Respiratory diseases cause considerable morbidity in autumn and winter and are a priority in public health monitoring. In Germany, they are subject to a number of surveillance systems, including both pathogen-specific and syndromic indicators. In this paper we present a collaborative multi-target and multi-model real-time forecasting system rolled out during the 2024/25 season, and discuss differences to earlier efforts carried out during the COVID-19 pandemic. A total of nine models were run to generate forecasts of general practitioner consultations for acute respiratory infections (ARI), hospitalizations for severe acute respiratory infections (SARI) and confirmed cases of seasonal influenza and RSV. As all indicators were subject to retrospective revisions, forecasting models were combined with a nowcasting step. Whenever multiple models were available for the same indicator, we combined them into an ensemble. Nowcasts showed convincing performance, even though for some models Christmas break effects led to an upward bias in early January. Forecasts were overall well-calibrated and most models outperformed simple benchmark models. These improvements were generally more substantial for age-stratified than pooled targets, and concentrated at lead times of two to three weeks. Anticipating the peak timing and magnitude proved to be challenging, with many models predicting too flat curves with a too early turnaround (e.g. already in late January rather than mid-February for SARI). The combined ensemble forecast was among the best-performing approaches, but unlike in previous related projects did not consistently outperform individual models. We conclude by discussing learnings on the organization of collaborative forecasting projects in post-COVID-19 times and the potential of AI-supported modelling.
- Managed Hermes Agent | Hostinger
Hermes Agent with zero setup, maintenance, or infrastructure management.
- AI Soundtrack
Generate custom copyright-free music that fits to your video automatically
- Genrank
The AI Search Platform for marketing teams.
- Mixelligence IA
The AI marketing team for freelancers and small businesses
- America needs to stop getting shocked by Chinese AI
Kimi K3 and Qwen3.8 should come as no surprise.
- Rerun
The easiest way to build AI agents for all your tasks
- OpenAI says Hugging Face was breached by its pre-release models
OpenAI has come forward to claim responsibility for the Hugging Face breach, saying it was the result of internal testing gone awry.
- OpenAI Says Its Models Accidentally Hacked Hugging Face
OpenAI said its advanced artificial intelligence models inadvertently hacked Hugging Face Inc. in an “unprecedented” incident that prompted fresh calls for curbs on the technology.
- OpenAI says AI models went rogue during testing, triggering 'unprecedented' breach at startup
OpenAI says AI models went rogue during testing, triggering 'unprecedented' breach at startup Reuters
- Hugging Face Breach Signals A New Era Of AI-Powered Cyberattacks
The Hugging Face breach suggests AI-powered cyberattacks are no longer theoretical. Experts explain what it means for defenders and what's coming next.
- OpenAI says its AI models escaped control and hacked into AI company Hugging Face
OpenAI says its AI models escaped control and hacked into AI company Hugging Face Fortune
- OpenAI Says Its AI Broke Containment, Went to Internet and Hacked Hugging Face
OpenAI Says Its AI Broke Containment, Went to Internet and Hacked Hugging Face The Information
- Hugging Face confirms data breach by AI agent: Why it has sparked a debate on cyber guardrails
Hugging Face confirms data breach by AI agent: Why it has sparked a debate on cyber guardrails
- OpenAI and Hugging Face partner to address security incident during model evaluation
OpenAI and Hugging Face share early findings from a security incident during AI model evaluation, highlighting advanced cyber capabilities and lessons for defenders.
- OpenAI Models Behind HuggingFace Cybersecurity Incident
Link: https://openai.com/index/hugging-face-model-evaluation-security-incident/ From the OpenAI blog post: Last week, Hugging Face disclosed a new kind of security incident after they detected and contained an AI agent that compromised their infrastructure , something we expect to become more commonplace with the proliferation of increasingly cyber-capable models. After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark of cyber capabilities. We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities , and are responding accordingly. We are sharing preliminary findings at this stage to help defenders understand what happened and to help calibrate on what models are now capable of. We will continue to conduct a thorough investigation alongside Hugging Face and will share more details on the vulnerabilities, incident, and findings when our investigation is complete. (emphasis added.) Yesterday, OpenAI disclosed that some of their internally models were misaligned. Today, they disclosed that "a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model" had compromised HuggingFace infrastructure in the course of running some OpenAI internal cyber evaluations on ExploitGym . These cyber evaluations were supposed to be run in sandboxed environments, with internet access limited to installing packages, then: The models identified a zero-day vulnerability to escape the sandbox, gained privileges in OpenAI research infrastructure, and eventually acquired open Internet access. The models reasoned that ExploitGym test solutions lived on HuggingFace servers. They then stole credentials and used zero-day vulnerabilities to hack HuggingFace's production infrastructure in order to get said solutions. HuggingFace noticed, contained the models, called law enforcement , fixed the vulnerabilities, and then disclosed this incident on July 16th. (The incident happened "earlier in the week"). The exact scope of the incident is unknown Discuss
- OpenAI: We (inadvertently) hacked Hugging Face (sorry)
The company didn't actually say "sorry."
- OpenAI Confirms Its AI Broke Out of a Sandbox and Breached Hugging Face
OpenAI said on Tuesday that two of its AI models, including the flagship Sol, broke out of a secure test environment, gained internet access by exploiting a zero-day vulnerability in third-party software, and hacked into Hugging Face’s production infrastructure. The company called the incident “unprecedented” and said it was sharing preliminary findings to help defenders […] This story continues at The Next Web
- OpenAI says its AI technology acted on its own in an 'unprecedented' hack of another company
OpenAI says its AI technology acted on its own in an 'unprecedented' hack of another company Toronto Star
- OpenAI says AI models went rogue during testing and hacked a startup
Company says models managed to reach the internet and compromise the infrastructure of Hugging Face
- OpenAI: Oops, Our Models Went Rogue, Hacked Hugging Face
OpenAI: Oops, Our Models Went Rogue, Hacked Hugging Face PCMag
- AI Platform Hugging Face Fends Off Hack From... AI
AI Platform Hugging Face Fends Off Hack From... AI PCMag
- OpenAI: Oops, Our Models Went Rogue, Hacked Hugging Face
OpenAI: Oops, Our Models Went Rogue, Hacked Hugging Face PCMag Australia
- OpenAI: Oops, Our Models Went Rogue, Hacked Hugging Face
OpenAI: Oops, Our Models Went Rogue, Hacked Hugging Face PCMag UK
- OpenAI says its own AI models broke out of testing and hacked Hugging Face
OpenAI Group PBC today disclosed that two of its artificial intelligence models broke out of a controlled testing environment and hacked open-source AI platform Hugging Face Inc. to cheat on an internal benchmark in what the company called an unprecedented cyber incident. The two models, OpenAI’s latest publicly available model GPT-5.6 Sol and a more capable […] The post OpenAI says its own AI models broke out of testing and hacked Hugging Face appeared first on SiliconANGLE .
- OpenAI says its AI technology acted on its own in an 'unprecedented' hack of another company
OpenAI has disclosed an "unprecedented cyber incident" where its AI system allegedly hacked into another AI company
- Agentic AI attack breaches Hugging Face
Agentic AI attack breaches Hugging Face Computing UK
- OpenAI says its AI technology acted on its own in an 'unprecedented' hack of another company
OpenAI says its AI technology acted on its own in an 'unprecedented' hack of another company San Francisco Chronicle
- OpenAI says its AI technology acted on its own in an 'unprecedented' hack of another company
OpenAI says its AI technology acted on its own in an 'unprecedented' hack of another company AP News
- OpenAI says it accidentally hacked Hugging Face with a new AI system
OpenAI says its AI models mistakenly breached open-source AI platform Hugging Face during internal testing. In a blog post on Tuesday, OpenAI writes that GPT-5.6 Sol and "an even more capable pre-release model" discovered vulnerabilities within their sandboxed testing environment, allowing them to gain access to the internet and target Hugging Face. On July 16th, […]
- OpenAI says its AI technology acted on its own in an ‘unprecedented’ hack of another company
OpenAI says its AI technology acted on its own in an ‘unprecedented’ hack of another company Boston Herald
- US judge approves Anthropic’s $1.5 billion settlement of copyright lawsuit
US judge approves Anthropic’s $1.5 billion settlement of copyright lawsuit
- Anthropic pays $1.5B to settle contentious copyright case
A US federal court has approved Anthropic’s $1.5 billion settlement in a class-action lawsuit in which authors accused the AI company of using their books without permission to train the AI model Claude. This is the largest such settlement to date in a US copyright case, according to Reuters . The dispute is one of several legal cases in which copyright holders sued AI companies over how large language models (LLMs) were trained, and it is the first major AI-related copyright dispute in the US to be resolved through a settlement. A judge had previously ruled that the actual training of AI models using books falls under the “fair use” doctrine in US copyright law. But Anthropic was found to have violated the law by storing more than 7 million pirated books in a central library, regardless of whether they were later used for AI training or not.
- Anthropic’s $1.5bn copyright lawsuit settlement gets approval
Each payout per work will amount to roughly $3,000, to be paid to around 500,000 works. Read more: Anthropic’s $1.5bn copyright lawsuit settlement gets approval
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot Toronto Star
- Judge Approves Anthropic's $1.5 Billion Copyright Settlement for Authors
Judge Approves Anthropic's $1.5 Billion Copyright Settlement for Authors PCMag Australia
- Judge Approves Anthropic's $1.5 Billion Copyright Settlement for Authors
Judge Approves Anthropic's $1.5 Billion Copyright Settlement for Authors PCMag UK
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot Chicago Tribune
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot San Francisco Chronicle
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot Houston Chronicle
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot AP News
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
A federal judge has approved a $1.5 billion copyright settlement involving AI company Anthropic
- Anthropic AI copyright lawsuit: Update as judge approves massive settlement and payout for authors
Anthropic will finally have to pay authors for pirating their books. On Monday, U.S. District Judge Araceli Martínez-Olguín approved a $1.5 billion AI copyright settlement . The payout is believed to be the largest ever for a U.S. copyright suit. Companies like Facebook parent Meta Platforms and ChatGPT maker OpenAI have also faced copyright lawsuits from authors. What does the Anthropic settlement say? Each member of the class action lawsuit could receive between $200 and $150,000 per work, according to the decision. However, it estimates that the per-work payment will be about $3,000. The payout stems from a 2024 copyright infringement lawsuit filed by a group of authors led by Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson. It claimed that Anthropic had committed “large-scale theft” in choosing to use pirated copies of books to train its chatbot, Claude. In June 2025, Judge William Alsup—now retired—ruled that AI could be trained on copyrighted books , but that Anthropic had accessed these books illegally, taking bootleg copies from online libraries. The case was set to go to trial in December and could have led to hundreds of billions of dollars in damages owed by Anthropic. Instead, the company chose to settle in September for $1.5 billion. Almost a year later, Judge Martínez-Olguín has taken over for Judge Alsup and approved the settlement. Fast Company reached out to Anthropic for comment. Monday was a busy day for Judge Martínez-Olguín, as she also granted the temporary halt of the Paramount-Warner Bros. merger. California led a 12-state coalition that argued the merger will violate antitrust law.
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot
Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot Boston Herald
- Bessent Says US to Scrutinize Chinese AI Models for IP Theft
Treasury Secretary Scott Bessent said the US will carefully examine open source artificial intelligence models from China for signs of intellectual property theft, amid concerns that a new wave of lower-cost Chinese competitors may sweep aside the top American models.
- Bessent says U.S. could sanction China over AI model 'theft'
Chinese open-weight models are gaining steam against leading offerings from American companies like OpenAI and Anthropic.
- Top American AI Execs Sound Alarm on Chinese Models
The White House is divided on how to respond to recent advances in Chinese AI and has weighed crackdown measures.