AWS Expands AI Model Access with Enhanced Security
In brief
- Amazon Web Services (AWS) has expanded access to Anthropic's Claude Fable 5 models through its Bedrock service.
- This move aims to provide customers with cutting-edge AI capabilities while ensuring robust security and privacy measures are in place.
- The models, known for their strong guardrails against misuse, are now available to AWS users, offering powerful tools for cybersecurity and other applications.
- The release of these advanced models follows extensive collaboration between AWS and Anthropic through Project Glasswing, which focuses on refining model guardrails to prevent adversaries from exploiting AI capabilities.
- This balance ensures that defenders can leverage the models' power without exposing potential vulnerabilities to malicious actors.
- Looking ahead, AWS emphasizes ongoing iteration with partners to enhance security and address emerging challenges in the rapidly evolving AI landscape.
- Users can expect continued updates as new models are released and guardrail systems evolve to meet industry needs.
Terms in this brief
- Bedrock
- A service by AWS that provides access to AI models for businesses and developers. It allows users to integrate advanced AI capabilities into their applications without needing deep technical expertise in machine learning or AI model management.
- Claude Fable 5
- One of Anthropic's AI models, known for its strong guardrails against misuse. Claude Fable 5 is designed to provide robust and secure AI capabilities, making it suitable for sensitive applications like cybersecurity.
Read full story at AWS ML Blog →
More briefs
ZLUDA Update Adds PhysX and Blender Support
ZLUDA has added two new major workloads: PhysX and Blender. This update also includes improved Windows support and better machine learning support. The new PhysX support means AMD GPU owners can play older games with higher frame rates and additional visual effects. For example, Mafia II can now run with maxed out settings and PhysX enabled on an AMD GPU. ZLUDA's updates will help gamers play classic games on newer hardware with better performance, and the project will continue to develop and improve.
AI Summaries Downplay Hotel Complaints
Tripadvisor's AI summaries of hotel reviews are downplaying serious complaints. A hotel being sued for food poisonings was described as spotless. A resort with guest complaints of sexual harassment by staff was praised for friendly service. Guests reported problems like raw chicken, flies on the buffet, and poor hygiene. The AI summaries are meant to help holidaymakers decide where to book. Tripadvisor will now look into the examples where reviews did not match the intended property.
BBVA Completes First AI-Initiated Transaction
BBVA completed a transaction initiated by an artificial intelligence agent on behalf of a cardholder. The payment was carried out using real card credentials and the systems of an active merchant. This transaction is important because it shows AI agents can securely complete purchases on behalf of cardholders within today's payments infrastructure. The transaction used Visa Payment Passkeys, a biometric authentication method that enables consumers to authorise online payments securely without passwords or SMS codes. 62% of surveyed consumers in Spain use AI to look for gift ideas, research products. The success of this transaction is a step forward for AI-enabled payments and BBVA will continue to explore how these new experiences can be delivered securely and at scale to cardholders. Next steps will be to further develop this technology to deliver seamless and reliable payment experiences.
AI Agents Now Build Reliable Web Scrapers Without Human Intervention
AI agents are now capable of building reliable web scrapers through a new framework that avoids common pitfalls like broken selectors and schema mismatches. This breakthrough comes from experiments on 138 tasks, where the system consistently produced accurate results by using a structured approach involving JSON configurations and six-type collector taxonomy. The key innovation lies in shifting LLM output from free-form code to typed JSON structures, which ensures deterministic and verifiable execution paths. On 80 independently verified tasks, this method achieved zero execution-stage errors and the lowest average wall-clock time, prioritizing reusable and reliable data collection over initial quality. This advancement marks a significant step toward more dependable AI-driven web scraping tools, enabling developers to streamline repetitive data extraction processes with fewer manual interventions. The framework's ability to handle repeated tasks efficiently could pave the way for broader adoption in industries reliant on real-time data collection.
AI Can Now Match Bookmakers' Accuracy in Predicting NFL Wins
Researchers have developed a new method for training AI to predict NFL game outcomes with unprecedented accuracy. By using a novel reward system based on past game data, the AI model achieved calibration levels matching professional betting markets without any human input or fine-tuning. This breakthrough addresses previous issues where AI predictions were less accurate due to noisy targets and corrupted reasoning chains. The innovation lies in a label-free reward mechanism that eliminates noise by estimating win rates from historical outcomes. By masking gradients during training, the AI maintains intact reasoning processes while avoiding corruption from policy gradients. This approach enables a 7B parameter model to reach betting market-level calibration through direct predictions alone. This development opens doors for more reliable probabilistic forecasting in various fields beyond sports analytics. Future research will focus on extending this method to other domains and improving generalization across different types of stochastic outcomes.