Events & Conferences
How Amazon’s Vulcan robots use touch to plan and execute motions
This week, at Amazon’s Delivering the Future symposium in Dortmund, Germany, Amazon announced that its Vulcan robots, which stow items into and pick items from fabric storage pods in Amazon fulfillment centers (FCs), have completed a pilot trial and are ready to move into beta testing.
Amazon FCs already use robotic arms to retrieve packages and products from conveyor belts and open-topped bins. But a fabric pod is more like a set of cubbyholes, accessible only from the front, and the items in the individual cubbies are randomly assorted and stacked and held in place by elastic bands. It’s nearly impossible to retrieve an item from a cubby or insert one into it without coming into physical contact with other items and the pod walls.
The Vulcan robots thus have end-of-arm tools — grippers or suction tools — equipped with sensors that measure force and torque along all six axes. Unlike the robot arms currently used in Amazon FCs, the Vulcan robots are designed to make contact with random objects in their work environments; the tool sensors enable them to gauge how much force they are exerting on those objects — and to back off before the force becomes excessive.
“A lot of traditional industrial automation — think of welding robots or even the other Amazon manipulation projects — are moving through free space, so the robot arms are either touching the top of a pile, or they’re not touching anything at all,” says Aaron Parness, a director of applied science with Amazon Robotics, who leads the Vulcan project. “Traditional industrial automation, going back to the ’90s, is built around preventing contact, and the robots operate using only vision and knowledge of where their joints are in space.
“What’s really new and unique and exciting is we are using a sense of touch in addition to vision. One of the examples I give is when you as a person pick up a coin off a table, you don’t command your fingers to go exactly to the specific point where you grab the coin. You actually touch the table first, and then you slide your fingers along the table until you contact the coin, and when you feel the coin, that’s your trigger to rotate the coin up into your grasp. You’re using contact both in the way you plan the motion and in the way you control the motion, and our robots are doing the same thing.”
The Vulcan pilot involved six Vulcan Stow robots in an FC in Spokane, Washington; the beta trial will involve another 30 robots in the same facility, to be followed by an even larger deployment at a facility in Germany, with Vulcan Stow and Vulcan Pick working together.
Vulcan Stow
The Vulcan Stow robot visualizes the volume of space necessary to stow a new item in a fabric pod, and to create that space, it uses its extensible blade to move other items to the side.
Inside the fulfillment center
When new items arrive at an FC, they are stowed in fabric pods at a stowing station; when a customer places an order, the corresponding items are picked from pods at a picking station. Autonomous robots carry the pods between the FC’s storage area and the stations. Picked items are sorted into totes and sent downstream for packaging.
The allocation of items to pods and pod shelves is fairly random. This may seem counterintuitive, but in fact it maximizes the efficiency of the picking and stowing operations. An FC might have 250 stowing stations and 100 picking stations. Random assortment minimizes the likelihood that any two picking or stowing stations will require the same pod at the same time.
To reach the top shelves of a pod, a human worker needs to climb a stepladder. The plan is for the Vulcan robots to handle the majority of stow and pick operations on the highest and lowest shelves, while humans will focus on the middle shelves and on more challenging operations involving densely packed bins or items, such as fluid containers, that require careful handling.
End-of-arm tools
The Vulcan robots’ main hardware innovation is the end-of-arm tools (EOATs) they use to perform their specialized tasks.
The pick robot’s EOAT is a suction device. It also has a depth camera to provide real-time feedback on the way in which the contents of the bin have shifted in response to the pick operation.
The stow EOAT is a gripper with two parallel plates that sandwich the item to be stowed. Each plate has a conveyer belt built in, and after the gripper moves into position, it remains stationary as the conveyer belts slide the item into position. The stow EOAT also has an extensible aluminum attachment that’s rather like a kitchen spatula, which it uses to move items in the bin aside to make space for the item being stowed.
Both the pick and stow robots have a second arm whose EOAT is a hook, which is used to pull down or push up the elastic bands covering the front of the storage bin.
The stow algorithm
As a prelude to the stow operation, the stow robot’s EOAT receives an item from a conveyor belt. The width of the gripper opening is based on a computer vision system’s inference of the item’s dimensions.
The stow system has three pairs of stereo cameras mounted on a tower, and their redundant stereo imaging allows it to build up a precise 3-D model of the pod and its contents.
At the beginning of a stow operation, the robot must identify a pod bin with enough space for the item to be stowed. A pod’s elastic bands can make imaging the items in each bin difficult, so the stow robot’s imaging algorithm was trained on synthetic bin images in which elastic bands were added by a generative-AI model.
The imaging algorithm uses three different deep-learning models to segment the bin image in three different ways: one model segments the elastic bands; one model segments the bins; and the third segments the objects inside the bands. These segments are then projected onto a three-dimensional point cloud captured by the stereo cameras to produce a composite 3-D segmentation of the bin.
The stow algorithm then computes bounding boxes indicating the free space in each bin. If the sum of the free-space measurements for a particular bin is adequate for the item to be stowed, the algorithm selects the bin for insertion. If the bounding boxes are non-contiguous, the stow robot will push items to the side to free up space.
The algorithm uses convolution to identify space in a 2-D image in which an item can be inserted: that is, it steps through the image applying the same kernel — which represents the space necessary for an insertion — to successive blocks of pixels until it finds a match. It then projects the convolved 2-D image onto the 3-D model, and a machine learning model generates a set of affordances indicating where the item can be inserted and, if necessary, where the EOAT’s extensible blade can be inserted to move objects in the bin to the side.
Based on the affordances, the stow algorithm then strings together a set of control primitives — such as approach, extend blade, sweep, and eject_item — to execute the stow. If necessary, the robot can insert the blade horizontally and rotate an object 90 degrees to clear space for an insertion.
“It’s not just about creating a world model,” Parness explains. “It’s not just about doing 3-D perception and saying, ‘Here’s where everything is.’ Because we’re interacting with the scene, we have to predict how that pile of objects will shift if we sweep them over to the side. And we have to think about like the physics of ‘If I collide with this T-shirt, is it going to be squishy, or is it going to be rigid?’ Or if I try and push on this bowling ball, am I going to have to use a lot of force? Versus a set of ping pong balls, where I’m not going to have to use a lot of force. That reasoning layer is also kind of unique.”
The pick algorithm
The first step in executing a pick operation is determining bin contents’ eligibility for robotic extraction: if a target object is obstructed by too many other objects in the bin, it’s passed to human pickers. The eligibility check is based on images captured by the FC’s existing imaging systems and augmented with metadata about the bins’ contents, which helps the imaging algorithm segment the bin contents.
The pick operation itself uses the EOAT’s built-in camera, which uses structured light — an infrared pattern projected across the objects in the camera’s field of view — to gauge depth. Like the stow operation, the pick operation begins by segmenting the image, but the segmentation is performed by a single MaskDINO neural model. Parness’s team, however, added an extra layer to the MaskDINO model, which classifies the segmented objects into four categories: (1) not an item (e.g., elastic bands or metal bars), (2) an item in good status (not obstructed), (3) an item below others, or (4) an item blocked by others.
Like the stow algorithm, the pick algorithm projects the segmented image onto a point cloud indicating the depths of objects in the scene. The algorithm also uses a signed distance function to characterize the three-dimensional scene: free space at the front of a bin is represented with positive distance values, and occupied space behind a segmented surface is represented with negative distance values.
Next — without scanning barcodes — the algorithm must identify the object to be picked. Since the products in Amazon’s catalogue are constantly changing, and the lighting conditions under which objects are imaged can vary widely, the object identification compares target images on the fly to sample product images captured during other FC operations.
The product-matching model is trained through contrastive learning: it’s fed pairs of images, either same product photographed from different angles and under different lighting conditions, or two different products; it learns to minimize the distance between representations of the same object in the representational space and to maximize the distance between representations of different objects. It thus becomes a general-purpose product matcher.
Using the 3-D composite, the algorithm identifies relatively flat surfaces of the target item that promise good adhesion points for the suction tool. Candidate surfaces are then ranked according to the signed distances of the regions around them, which indicate the likelihood of collisions during extraction.
Finally, the suction tool is deployed to affix itself to the highest-ranked candidate surface. During the extraction procedure, the suction pressure is monitored to ensure a secure hold, and the camera captures 10 low-res images per second to ensure that the extraction procedure hasn’t changed the geometry of the bin. If the initial pick point fails, the robot tries one of the other highly ranked candidates. In the event of too many failures, it passes the object on for human extraction.
“I really think of this as a new paradigm for robotic manipulation,” Parness says. “Getting out of the ‘I can only move through free space’ or ‘Touch the thing that’s on the top of the pile’ to the new paradigm where I can handle all different kinds of items, and I can dig around and find the toy that’s at the bottom of the toy chest, or I can handle groceries and pack groceries that are fragile in a bag. I think there’s maybe 20 years of applications for this force-in-the-loop, high-contact style of manipulation.”
For more information about the Vulcan Pick and Stow robots, see the associated research papers: Pick | Stow.
Have you ever worked is legacy code? Are you curious what it takes to modernize systems at a massive scale?
Pascal Hartig is joined on the latest Meta Tech Podcast by Elaine and Buping, two software engineers working on a bold project to rewrite the decades-old C code in one of Meta’s core messaging libraries in Rust. It’s an ambitious effort that will transform a central messaging library that is shared across Messenger, Facebook, Instagram, and Meta’s AR/VR platforms.
They discuss taking on a project of this scope – even without a background in Rust, how they’re approaching it, and what it means to optimize for ‘developer happiness.’
Download or listen to the episode below:
You can also find the episode wherever you get your podcasts, including:
The Meta Tech Podcast is a podcast, brought to you by Meta, where we highlight the work Meta’s engineers are doing at every level – from low-level frameworks to end-user features.
Send us feedback on Instagram, Threads, or X.
And if you’re interested in learning more about career opportunities at Meta visit the Meta Careers page.
Amazon Research Awards (ARA) provides unrestricted funds and AWS Promotional Credits to academic researchers investigating various research topics in multiple disciplines. This cycle, ARA received many excellent research proposals from across the world and today is publicly announcing 73 award recipients who represent 46 universities in 10 countries.
This announcement includes awards funded under five call for proposals during the fall 2024 cycle: AI for Information Security, Automated Reasoning, AWS AI, AWS Cryptography, and Sustainability. Proposals were reviewed for the quality of their scientific content and their potential to impact both the research community and society. Additionally, Amazon encourages the publication of research results, presentations of research at Amazon offices worldwide, and the release of related code under open-source licenses.
Recipients have access to more than 700 Amazon public datasets and can utilize AWS AI/ML services and tools through their AWS Promotional Credits. Recipients also are assigned an Amazon research contact who offers consultation and advice, along with opportunities to participate in Amazon events and training sessions.
“Automated Reasoning is an important area of research for Amazon, with potential applications across various features and applications to help improve security, reliability, and performance for our customers. Through the ARA program, we collaborate with leading academic researchers to explore challenges in this field,” said Robert Jones, senior principal scientist with the Cloud Automated Reasoning Group. “We were again impressed by the exceptional response to our Automated Reasoning call for proposals this year, receiving numerous high-quality submissions. Congratulations to the recipients! We’re excited to support their work and partner with them as they develop new science and technology in this important area.”
“At Amazon, we believe that solving the world’s toughest sustainability challenges benefits from both breakthrough scientific research and open and bold collaboration. Through programs like the Amazon Research Awards program, we aim to support academic research that could contribute to our understanding of these complex issues,” said Kommy Weldemariam, Director of Science and Innovation Sustainability. “The selected proposals represent innovative projects that we hope will help advance knowledge in this field, potentially benefiting customers, communities, and the environment.”
ARA funds proposals throughout the year in a variety of research areas. Applicants are encouraged to visit the ARA call for proposals page for more information or send an email to be notified of future open calls.
The tables below list, in alphabetical order by last name, fall 2024 cycle call-for-proposal recipients, sorted by research area.
AI for Information Security
| Recipient | University | Research title |
| Christopher Amato | Northeastern University | Multi-Agent Reinforcement Learning Cyber Defense for Securing Cloud Computing Platforms |
| Bernd Bischl | Ludwig Maximilian University of Munich | Improving Generative and Foundation Models Reliability via Uncertainty-awareness |
| Shiqing Ma | University Of Massachusetts Amherst | LLM and Domain Adaptation for Attack Detection |
| Alina Oprea | Northeastern University | Multi-Agent Reinforcement Learning Cyber Defense for Securing Cloud Computing Platforms |
| Roberto Perdisci | University of Georgia | ContextADBench: A Comprehensive Benchmark Suite for Contextual Anomaly Detection |
Automated Reasoning
| Recipient | University | Research title |
| Nada Amin | Harvard University | LLM-Augmented Semi-Automated Proofs for Interactive Verification |
| Suguman Bansal | Georgia Institute of Technology | Certified Inductive Generalization in Reinforcement Learning |
| Ioana Boureanu | University of Surrey | Phoebe+: An Automated-Reasoning Tool for Provable Privacy in Cryptographic Systems |
| Omar Haider Chowdhury | Stony Brook University | Restricter: An Automatic Tool for Authoring Amazon Cedar Access Control Policies with the Principle of Least Privilege |
| Stefan Ciobaca | Alexandru Ioan Cuza University | An Interactive Proof Mode for Dafny |
| João Ferreira | INESC-ID | Polyglot Automated Program Repair for Infrastructure as Code |
| Sicun Gao | University Of California, San Diego | Monte Carlo Trees with Conflict Models for Proof Search |
| Mirco Giacobbe | University of Birmingham | Neural Software Verification |
| Tobias Grosser | University of Cambridge | Synthesis-based Symbolic BitVector Simplification for Lean |
| Ronghui Gu | Columbia University | Scaling Formal Verification of Security Properties for Unmodified System Software |
| Alexey Ignatiev | Monash University | Huub: Next-Gen Lazy Clause Generation |
| Kenneth McMillan | University of Texas At Austin | Synthesis of Auxiliary Variables and Invariants for Distributed Protocol Verification |
| Alexandra Mendes | University of Porto | Overcoming Barriers to the Adoption of Verification-Aware Languages |
| Jason Nieh | Columbia University | Scaling Formal Verification of Security Properties for Unmodified System Software |
| Rohan Padhye | Carnegie Mellon University | Automated Synthesis and Evaluation of Property-Based Tests |
| Nadia Polikarpova | University Of California, San Diego | Discovering and Proving Critical System Properties with LLMs |
| Fortunat Rajaona | University of Surrey | Phoebe+: An Automated-Reasoning Tool for Provable Privacy in Cryptographic Systems |
| Subhajit Roy | Indian Institute of Technology Kanpur | Theorem Proving Modulo LLM |
| Gagandeep Singh | University of Illinois At Urbana–Champaign | Trustworthy LLM Systems using Formal Contracts |
| Scott Stoller | Stony Brook University | Restricter: An Automatic Tool for Authoring Amazon Cedar Access Control Policies with the Principle of Least Privilege |
| Peter Stuckey | Monash University | Huub: Next-Gen Lazy Clause Generation |
| Yulei Sui | University of New South Wales | Path-Sensitive Typestate Analysis through Sparse Abstract Execution |
| Nikos Vasilakis | Brown University | Semantics-Driven Static Analysis for the Unix/Linux Shell |
| Ping Wang | Stevens Institute of Technology | Leveraging Large Language Models for Reasoning Augmented Searching on Domain-specific NoSQL Database |
| John Wawrzynek | University of California, Berkeley | GPU-Accelerated High-Throughput SAT Sampling |
AWS AI
| Recipient | University | Research title |
| Panagiotis Adamopoulos | Emory University | Generative AI solutions for The Spillover Effect of Fraudulent Reviews on Product Recommendations |
| Vikram Adve | University of Illinois at Urbana–Champaign | Fellini: Differentiable ML Compiler for Full-Graph Optimization for LLM Models |
| Frances Arnold | California Institute of Technology | Closed-loop Generative Machine Learning for De Novo Enzyme Discovery and Optimization |
| Yonatan Bisk | Carnegie Mellon University | Useful, Safe, and Robust Multiturn Interactions with LLMs |
| Shiyu Chang | University of California, Santa Barbara | Cut the Crap: Advancing the Efficient Communication of Multi-Agent Systems via Spatial-Temporal Topology Design and KV Cache Sharing |
| Yuxin Chen | University of Pennsylvania | Provable Acceleration of Diffusion Models for Modern Generative AI |
| Tianlong Chen | University of North Carolina at Chapel Hill | Cut the Crap: Advancing the Efficient Communication of Multi-Agent Systems via Spatial-Temporal Topology Design and KV Cache Sharing |
| Mingyu Ding | University of North Carolina at Chapel Hill | Aligning Long Videos and Language as Long-Horizon World Models |
| Nikhil Garg | Cornell University | Market Design for Responsible Multi-agent LLMs |
| Jessica Hullman | Northwestern University | Human-Aligned Uncertainty Quantification in High Dimensions |
| Christopher Jermaine | Rice University | Fast, Trusted AI Using the EINSUMMABLE Compiler |
| Yunzhu Li | Columbia University | Physics-Informed Foundation Models Through Embodied Interactions |
| Pattie Maes | Massachusetts Institute of Technology | Understanding How LLM Agents Deviate from Human Choices |
| Sasa Misailovic | University of Illinois at Urbana–Champaign | Fellini: Differentiable ML Compiler for Full-Graph Optimization for LLM Models |
| Kristina Monakhova | Cornell University | Trustworthy extreme imaging for science using interpretable uncertainty quantification |
| Todd Mowry | Carnegie Mellon University | Efficient LLM Serving on Trainium via Kernel Generation |
| Min-hwan Oh | Seoul National University | Mutually Beneficial Interplay Between Selection Fairness and Context Diversity in Contextual Bandits |
| Patrick Rebeschini | University of Oxford | Optimal Regularization for LLM Alignment |
| Jose Renau | University of California, Santa Cruz | Verification Constrained Hardware Optimization using Intelligent Design Agentic Programming |
| Vilma Todri | Emory University | Generative AI solutions for The Spillover Effect of Fraudulent Reviews on Product Recommendations |
| Aravindan Vijayaraghavan | Northwestern University | Human-Aligned Uncertainty Quantification in High Dimensions |
| Wei Yang | University of Texas at Dallas | Optimizing RISC-V Compilers with RISC-LLM and Syntax Parsing |
| Huaxiu Yao | University of North Carolina at Chapel Hill | Aligning Long Videos and Language as Long-Horizon World Models |
| Amy Zhang | University of Washington | Tools for Governing AI Agent Autonomy |
| Ruqi Zhang | Purdue University | Efficient Test-time Alignment for Large Language Models and Large Multimodal Models |
| Zheng Zhang | Rutgers University-New Brunswick | AlphaQC: An AI-powered Quantum Circuit Optimizer and Denoiser |
AWS Cryptography
| Recipient | University | Research title |
| Alexandra Boldyreva | Georgia Institute of Technology | Quantifying Information Leakage in Searchable Encryption Protocols |
| Maria Eichlseder | Graz University of Technology, Austria | SALAD – Systematic Analysis of Lightweight Ascon-based Designs |
| Venkatesan Guruswami | University of California, Berkeley | Obfuscation, Proof Systems, and Secure Computation: A Research Program on Cryptography at the Simons Institute for the Theory of Computing |
| Joseph Jaeger | Georgia Institute of Technology | Analyzing Chat Encryption for Group Messaging |
| Aayush Jain | Carnegie Mellon | Large Scale Multiparty Silent Preprocessing for MPC from LPN |
| Huijia Lin | University of Washington | Large Scale Multiparty Silent Preprocessing for MPC from LPN |
| Hamed Nemati | KTH Royal Institute of Technology | Trustworthy Automatic Verification of Side-Channel Countermeasures for Binary Cryptographic Programs using the HoIBA libary |
| Karl Palmskog | KTH Royal Institute of Technology | Trustworthy Automatic Verification of Side-Channel Countermeasures for Binary Cryptographic Programs using the HoIBA libary |
| Chris Peikert | University of Michigan, Ann Arbor | Practical Third-Generation FHE and Bootstrapping |
| Dimitrios Skarlatos | Carnegie Mellon University | Scale-Out FHE LLMs on GPUs |
| Vinod Vaikuntanathan | Massachusetts Institute of Technology | Can Quantum Computers (Really) Factor? |
| Daniel Wichs | Northeastern University | Obfuscation, Proof Systems, and Secure Computation: A Research Program on Cryptography at the Simons Institute for the Theory of Computing |
| David Wu | University Of Texas At Austin | Fast Private Information Retrieval and More using Homomorphic Encryption |
Sustainability
| Recipient | University | Research title |
| Meeyoung Cha | Max Planck Institute | Forest-Blossom (Flossom): A New Framework for Sustaining Forest Biodiversity Through Outcome-Driven Remote Sensing Monitoring |
| Jingrui He | University of Illinois at Urbana–Champaign | Foundation Model Enabled Earth’s Ecosystem Monitoring |
| Pedro Lopes | University of Chicago | AI-powered Tools that Enable Engineers to Make & Re-make Sustainable Hardware |
| Cheng Yaw Low | Max Planck Institute | Forest-Blossom (Flossom): A New Framework for Sustaining Forest Biodiversity Through Outcome-Driven Remote Sensing Monitoring |
AI safety is a priority at Amazon. Our investment in safe, transparent, and responsible AI (RAI) includes collaboration with the global community and policymakers. We are members of and collaborate with organizations such as the Frontier Model Forum, the Partnership on AI, and other forums organized by government agencies such as the National Institute of Standards and Technology (NIST). Consistent with Amazon’s endorsement of the Korea Frontier AI Safety Commitments, we published our Frontier Model Safety Framework earlier this year.
During the development of the Nova Premier model, we conducted a comprehensive evaluation to assess its performance and safety. This included testing on both internal and public benchmarks and internal/automated and third-party red-teaming exercises. Once the final model was ready, we prioritized obtaining unbiased, third-party evaluations of the model’s robustness against RAI controls. In this post, we outline the key findings from these evaluations, demonstrating the strength of our testing approach and Amazon Premier’s standing as a safe model. Specifically, we cover our evaluations with two third-party evaluators: PRISM AI and ActiveFence.
Evaluation of Nova Premier against PRISM AI
PRISM Eval’s Behavior Elicitation Tool (BET) dynamically and systematically stress-tests AI models’ safety guardrails. The methodology focuses on measuring how many adversarial attempts (steps) it takes to get a model to generate harmful content across several key risk dimensions. The central metric is “steps to elicit” — the number of increasingly sophisticated prompting attempts required before a model generates an inappropriate response. A higher number of steps indicates stronger safety measures, as the model is more resistant to manipulation. The PRISM risk dimensions (inspired by the MLCommons AI Safety Benchmarks) include CBRNE weapons, violent crimes, non-violent crimes, defamation, and hate, amongst several others.
Using the BET Eval tool and its V1.0 metric, which is tailored toward non-reasoning models, we compared the recently released Nova models (Pro and Premier) to the latest models in the same class: Claude (3.5 v2 and 3.7 non-reasoning) and Llama4 Maverick, all available through Amazon Bedrock. PRISM BET conducts black-box evaluations (where model developers don’t have access to the test prompts) of models integrated with their API. The evaluation conducted with BET Eval MAX, PRISM’s most comprehensive/aggressive testing suite, revealed significant variations in safety against malicious instructions. Nova models demonstrated superior overall safety performance, with an average of 43 steps for Premier and 52 steps for Pro, compared to 37.7 for Claude 3.5 v2 and fewer than 12 steps for other models in the comparison set (namely, 9.9 for Claude3.7, 11.5 for Claude 3.7 thinking, and 6.5 for Maverick). This higher step count suggests that on average, Nova’s safety guardrails are more sophisticated and harder to circumvent through adversarial prompting. The figure below presents the number of steps per harm category evaluated through BET Eval MAX.
The PRISM evaluation provides valuable insights into the relative safety of different Amazon Bedrock models. Nova’s strong performance, particularly in hate speech and defamation resistance, represents meaningful progress in AI safety. However, the results also highlight the ongoing challenge of building truly robust safety measures into AI systems. As the field continues to evolve, frameworks like BET will play an increasingly important role in benchmarking and improving AI safety. As a part of this collaboration Nicolas Miailhe, CEO of PRISM Eval, said, “It’s incredibly rewarding for us to see Nova outperforming strong baselines using the BET Eval MAX; our aim is to build a long-term partnership toward safer-by-design models and to make BET available to various model providers.” Organizations deploying AI systems should carefully consider these safety metrics when selecting models for their applications.
Manual red teaming with ActiveFence
The AI safety & security company ActiveFence benchmarked Nova Premier on Bedrock on prompts distributed across Amazon’s eight core RAI categories. ActiveFence also evaluated Claude 3.7 (non-reasoning mode) and GPT 4.1 API on the same set. The flag rate on Nova Premier was lower than that on the other two models, indicating that Nova Premier is the safest of the three.
| Model | 3P Flag Rate [↓ is better] |
| Nova Premier | 12.0% |
| Sonnet 3.7 (non-reasoning) | 20.6% |
| GPT4.1 API | 22.4% |
“Our role is to think like an adversary but act in service of safety,” said Guy Paltieli from ActiveFence. “By conducting a blind stress test of Nova Premier under realistic threat scenarios, we helped evaluate its security posture in support of Amazon’s broader responsible-AI goals, ensuring the model could be deployed with greater confidence.”
These evaluations conducted with PRISM and ActiveFence give us confidence in the strength of our guardrails and our ability to protect our customers’ safety when they use our models. While these evaluations demonstrate strong safety performance, we recognize that AI safety is an ongoing challenge requiring continuous improvement. These assessments represent a point-in-time snapshot, and we remain committed to regular testing and enhancement of our safety measures. No AI system can guarantee perfect safety in all scenarios, which is why we maintain monitoring and response systems after deployment.
Acknowledgments: Vincent Ponzo, Elyssa Vincent
-
Funding & Business6 days agoKayak and Expedia race to build AI travel agents that turn social posts into itineraries
-
Jobs & Careers6 days ago
Mumbai-based Perplexity Alternative Has 60k+ Users Without Funding
-
Mergers & Acquisitions6 days ago
Donald Trump suggests US government review subsidies to Elon Musk’s companies
-
Funding & Business6 days ago
Rethinking Venture Capital’s Talent Pipeline
-
Jobs & Careers6 days ago
Why Agentic AI Isn’t Pure Hype (And What Skeptics Aren’t Seeing Yet)
-
Funding & Business3 days ago
Sakana AI’s TreeQuest: Deploy multi-model teams that outperform individual LLMs by 30%
-
Funding & Business6 days agoFrom chatbots to collaborators: How AI agents are reshaping enterprise work
-
Jobs & Careers3 days ago
Ilya Sutskever Takes Over as CEO of Safe Superintelligence After Daniel Gross’s Exit
-
Funding & Business3 days ago
Dust hits $6M ARR helping enterprises build AI agents that actually do stuff instead of just talking
-
Jobs & Careers6 days ago
Telangana Launches TGDeX—India’s First State‑Led AI Public Infrastructure
