Events & Conferences
How Amazon Robotics researchers are solving a “beautiful problem”
The rate of innovation in machine learning is simply off the chart — what is possible today was barely on the drawing board even a handful of years ago. At Amazon, this has manifested in a robotic system that can not only identify potential space in a cluttered storage bin, but also sensitively manipulate that bin’s contents to create that space before successfully placing additional items inside — a result that, until recently, was impossible.
This journey starts when a product arrives at an Amazon fulfillment center (FC). The first order of business is to make it available to customers by adding it to the FC’s available inventory.
In practice, this means picking it up and stowing it in a storage pod. A pod is akin to a big bookcase, made of sturdy yellow fabric, that comprises up to 40 cubbies, known as bins. Each bin has strips of elastic across its front to keep the items inside from falling out. These pods are carried by a wheeled robot, or drive unit, to the workstation of the Amazon associate doing the stowing. When the pod is mostly full, it is wheeled back into the warehouse, where the items it contains await a customer order.
Stowing is a major component of Amazon’s operations. It is also a task that seemed an intractable problem from a robotic automation perspective, due to the subtlety of thought and dexterity required to do the job.
Picture the task. You have an item for stowing in your hand. You gauge its size and weight. You look at the array of bins before you, implicitly perceiving which are empty, which are already full, which bins have big chunks of space in them, and which have the potential to make space if you, say, pushed all the items currently in the bin to one side. You select a bin, move the elastic out of the way, make room for the item, and pop it in. Job done. Now repeat.
“Breaking all existing industrial robot thinking”
This stow task requires two high-level capabilities not generally found in robots. One, an excellent understanding of the three-dimensional world. Two, the ability to manipulate a wide range of packaged but sometimes fragile objects — from lightbulbs to toys — firmly, but sensitively: pushing items gently aside, flipping them up, slotting one item at an angle between other items and so on.
A simulation of robotic stowing
For a robotic system to stand a chance at this task, it would need intelligent visual perception, a free-moving robot arm, an end-of-arm manipulator unknown to engineering, and a keen sense of how much force it is exerting. In short: good luck with that.
“Stow fundamentally breaks all existing industrial robotic thinking,” says Siddhartha Srinivasa, director of Amazon Robotics AI. “Industrial manipulators are typically bulky arms that execute fixed trajectories very precisely. It’s very positional.”
When Srinivasa joined Amazon in 2018, multiple robotics programs had already attempted to stow to fabric pods using stiff positional manipulators.
“They failed miserably at it because it’s a nightmare. It just doesn’t work unless you have the right computational tool: you must not think physically, but computationally.”
Srinivasa knew the science for robotic stow didn’t exist yet, but he knew the right people to hire to develop it. He approached Parker Owan as he completed his PhD at the University of Washington.
A “beautiful problem”
“At the time I was working on robotic contact, imitation learning, and force control,” says Owan, now a Robotics AI senior applied scientist. “Sidd said ‘Hey, there’s this beautiful problem at Amazon that you might be interested in taking a look at’, and he left it at that.”
The seed was planted. Owan joined Amazon, and then in 2019 dedicated himself to the stow challenge.
“I came at it from the perspective of decision-making algorithms: the perception needs; how to match items to the appropriate bin; how to leverage information of what’s in the bin to make better decisions; motion planning for a robot arm moving through free space; and then actually making contact with products and creating space in bins.”
About six months into his exploratory work, Owan was joined by a small team of applied scientists, and hardware expert Aaron Parness, now a Robotics AI senior manager of applied science. Parness admits he was skeptical.
“My initial reaction was ‘Oh, how brave and naïve that this guy, fresh out of his PhD, thinks robots can deal with this level of clutter and physical contact!’”
But Parness was quickly hooked. “Once you see how the problem can be broken down and structured, it suddenly becomes clear that there’s something super useful and interesting here.”
“Uncharted territory”
From a hardware perspective, the team needed to find a robot arm with force feedback. They tried several, before the team landed on an effective model. The arm provides feedback hundreds of times per second on how much force it is applying and any resistance it is meeting. Using this information to control the robot is called compliant manipulation.
“We knew from the beginning that we needed compliant manipulation, and we hadn’t seen anybody in industry do this at scale before,” says Owan. “It was uncharted territory.”
Parness got to work on the all-important hardware. The problem of moving the elastics aside to stow an item was resolved using a relatively simple hooking system.
How the band separator works
The end-of-arm tool (EOAT) proved to be a next-level challenge. One reason that stowing is difficult for robots is the sheer diversity of items Amazon sells, and their associated packaging. You might have an unpumped soccer ball next to a book, next to a sports drink, next to a T-shirt, next to a jewelry box. A robot would need to handle this level of variety. The EOAT evolved quickly over two years, with multiple failures and iterations.
Paddles grip an array of items
“In the end, we found that gently squeezing an item between two paddles was the more stable way to hold items than using suction cups or mechanical pinchers,” says Parness.
However, the paddle set up presented a challenge when trying to insert held items into bins — the paddles kept getting in the way. Parness and his growing team hit upon an alternative: holding the item next to a bin, before simultaneously opening the paddles and using a plunger to push the item in. This drop-and-push technique was prone to errors because not all items reacted to it in the same way.
The EOAT’s next iteration saw the team put miniature conveyor belts on each paddle, enabling the EOAT to feed items smoothly into the bins without having to enter the bin itself.
The miniature conveyor belt works to bring an item to its designated bin
“With that change, our stowing success rate jumped from about 80% to 99%. That was a eureka moment for us — we knew we had our winner,” says Parness.
Making space with motion primitives
The ability to place items in bins is crucial, but so is making space in cluttered bins. To better understand what would be required of the robot system, the team closely studied how they performed the task themselves. Owan even donned a head camera to record his efforts.
The team was surprised to find that the vast majority of space-making hand movements within a fabric bin could be boiled down to four types or “motion primitives”. These include a sideways sweep of the bin’s current contents, flipping upright things that are lying flat, stacking, and slotting something at an angle into the gap between other items.
The process of making space
The engineers realized that the EOAT’s paddles could not get involved with this bin-manipulation task, because they would get in the way. The solution, in the end, was surprisingly simple: a thin metal sheet that could extend from the EOAT, dubbed “the spatula”. The extended spatula can firmly, but sensitively, push items to one side, flip them up, and generally be used to make room in a bin, before the paddles eject an item into the space created.
But how does the system know how full the pod’s bins are, and how does it decide where, and how, it will make space for the next item to be stowed? This is where visual perception and machine learning come into play.
Deciding where to attempt to stow an item requires a good understanding of how much space, in total, is available in each fabric bin. In an ideal world, this is where 3D sensor technologies such as LiDAR would be used. However, because the elastic cords across the front of every bin partially blocks the view inside, this option isn’t feasible.
A robot arm executes motion primitives
Instead, the system’s visual perception is based on cameras pointed at the pod that feed their image data to a machine learning system. Based on what it can see of each bin’s contents, the system “erases” the elastics and models what is lying unseen in the bin, and then estimates the total available space in each of the pod’s bins.
Often there is space available in a cluttered bin, but it is not contiguous: there are pockets of space here and there. The ML system — based in part on existing models developed by the Amazon Fulfillment Technologies team — then predicts how much contiguous space it can create in each bin, given the motion primitives at its disposal.
How the perception system “sees” available space
“These primitives, each of which can be varied as needed, can be chained in infinitely many ways,” Srinivasa explains. “It can, say, flip it over here, then push it across and drop the item in. Humans are great at identifying these primitives in the first place, and machine learning is great at organizing and orchestrating them.”
When the system has a firm idea of the options, it considers the items in its buffer — an area near the robot arm’s gantry in which products of various shapes and sizes wait to be stowed — and decides which items are best placed in which bins for maximum efficiency.
“For every potential stow, the system will predict its likelihood of success,” says Parness. “When the best prediction of success falls to about 96%, which happens when a pod is nearly full, we send that pod off and wheel in a new one.”
“Robots and people work together”
At the end of summer 2021, with its potential feasibility and value becoming clearer, the senior leadership team at Amazon gave the project their full backing.
“They said ‘As fast as you can go; whatever you need’. So this year has been a wild, wild ride. It feels like we’re a start-up within Amazon,” says Parness, who noted the approach has significant advantages for FC employees as well.
“Robots and people work together in a hybrid system. Robots handle repetitive tasks and easily reach to the high and low shelves. Humans handle more complex items that require intuition and dexterity. The net effect will be more efficient operations that are also safer for our workers.”
Prototypes of the robotic stow workstation are installed at a lab in Seattle, Washington, and another system has been installed at an FC in Sumner, Washington, where it deals with live inventory. Already, the prototypes are stowing items well and showcasing the viability of the system.
“And there are always four or five scientists and engineers hovering around the robot, documenting issues and looking for improvements,” says Parness.
Stow will be the first brownfield automation project, at scale, at Amazon. We’re enacting a future in which robots and humans can actually work side by side without us having to dramatically change the human working environment.
This year, in a stowing test designed to include a variety of challenging product attributes — bagged items, irregular items with an offset center of gravity, and so on — the system successfully stowed 94 of 95 items. Of course, some items can never be stowed by this system, including particularly bulky or heavy products, or cylindrical items that don’t behave themselves on conveyor belts. The team’s ultimate target is to be able to stow 85% of products stocked by a standard Amazon FC.
“Interacting with chaotic arrangements of items, unknown items with different shapes and sizes, and learning to manipulate them in intelligent ways, all at Amazon scale — this is ground-breaking,” says Owan. “I feel like I’m at ground zero for a big thing, and that’s what makes me excited to come to work every day.”
“Stow will be the first brownfield automation project, at scale, at Amazon,” says Srinivasa. “Surgically inserting automation into existing buildings is very challenging, but we’re enacting a future in which robots and humans can actually work side by side without us having to dramatically change the human working environment.
“One of the advantages of the type of brownfield automation we do at Robotics AI is that it’s minimally disruptive to the process flow or the building space, which means that our robots can truly work alongside humans,” Srinivasa adds. “This is also a future benefit of compliant arms as they can, via software and AI, be made safer than industrial arms.”
Robots and humans working side by side is key to the long-term expansion of this technology beyond retail, says Parness.
“Think of robots loading delicate groceries or, longer term, loading dishwashers or helping people with tasks around the house. Robots with a sense of force in their control loop is a new paradigm in compliant-robotics applications.”
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 & Business1 week agoKayak and Expedia race to build AI travel agents that turn social posts into itineraries
-
Jobs & Careers1 week ago
Mumbai-based Perplexity Alternative Has 60k+ Users Without Funding
-
Mergers & Acquisitions1 week ago
Donald Trump suggests US government review subsidies to Elon Musk’s companies
-
Funding & Business7 days ago
Rethinking Venture Capital’s Talent Pipeline
-
Jobs & Careers7 days ago
Why Agentic AI Isn’t Pure Hype (And What Skeptics Aren’t Seeing Yet)
-
Funding & Business4 days ago
Sakana AI’s TreeQuest: Deploy multi-model teams that outperform individual LLMs by 30%
-
Jobs & Careers7 days ago
Astrophel Aerospace Raises ₹6.84 Crore to Build Reusable Launch Vehicle
-
Jobs & Careers7 days ago
Telangana Launches TGDeX—India’s First State‑Led AI Public Infrastructure
-
Funding & Business1 week agoFrom chatbots to collaborators: How AI agents are reshaping enterprise work
-
Jobs & Careers5 days ago
Ilya Sutskever Takes Over as CEO of Safe Superintelligence After Daniel Gross’s Exit
