NUWA · RESEARCH
Publications, technical reports, and research insights on frontier AI risk, agent safety, and system security.
NUWA · RESEARCH HIGHLIGHT
Starting with limited target information, AgentCyberRange evaluates how agents discover attack surfaces, exploit vulnerabilities, and carry access, credentials, and network reachability through a multistage attack.
Access requests from 100+ research institutions worldwide
89 results
Presents AgentDoS, a lifecycle-aware fuzzing framework for detecting resource-abuse DoS vulnerabilities in LLM-based agents.
Figure 3 · PDF p.5Uses simulation-derived reasoning correction to reduce unsafe actions in computer-use agents while preserving task utility.
Figure 1 · PDF p.2Introduces Thought-Aligner, a plug-in method that causally corrects unsafe agent thoughts before actions are executed.
Figure 1 · PDF p.2Synthesizes executable risk-evaluation environments that combine deterministic code state with LLM-generated narrative dynamics.
Figure 2 · PDF p.3Studies security risks and defenses for agents and foundation models.
Uncovers insecure resource-management practices in app-in-app cloud services.
Figure 1 · PDF p.1Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Detects taint-style vulnerabilities across C and Lua boundaries in firmware web services.
Figure 4 · PDF p.7Detects unsafe diffusion outputs during the generation process by approximating latent decoding, enabling earlier and cheaper NSFW intervention.
Figure 2 · PDF p.4Rehosts MCU firmware on KVM with near-native execution for scalable security analysis.
Figure 2 · PDF p.5
Reveals identity-confusion risks in email alias ecosystems and measures their security impact.
Figure 1 · PDF p.2Demonstrates autonomous agents acquiring external computational resources and propagating across remote devices under controlled, simulated real-world conditions.
Figure 2 · PDF p.3
Builds a lightweight framework to evaluate deception risk and user trust dynamics in open-ended human-AI dialogue.
Figure 1 · PDF p.3Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Frames deception targeting developers as a distinct frontier-AI risk and proposes recommendations for monitorability, evaluation integrity, and non-evadable control.
Figure 1 · PDF p.2Evaluates LLM criminal potential across traits such as false statements, framing, psychological manipulation, emotional disguise, and moral disengagement.
Figure 1 · PDF p.2Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Proposes an evolutionary prompt-injection attack that targets black-box LLM-powered tabular agents under structural payload constraints.
Figure 1 · PDF p.1Measures how LLM-enhanced search systems respond to adversarial SEO manipulation.
Figure 2 · PDF p.3Characterizes aggressive advertising behavior and user risk across mini-game ecosystems.
Figure 1 · PDF p.1
Extends self-replication evaluation across 32 AI systems and reports autonomous replication, self-exfiltration, adaptation, and shutdown-survival behaviors.
Figure 1 · PDF p.3Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Figure 2 · PDF p.3Identifies remote software versions through functional changes instead of exploit-only probes.
Figure 2 · PDF p.5Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Figure 2 · PDF p.6Demonstrates invisible HTML techniques that mislead both recipients and spam detectors.
Figure 1 · PDF p.2
Studies whether models recognize evaluation contexts and alter behavior, identifying observer effects that threaten safety-evaluation integrity.
Figure 1 · PDF p.3Evaluates whether frontier AI systems can autonomously self-replicate and reports successful self-replication in controlled trials.
Figure 1 · PDF p.4Uses service-aware feedback to improve vulnerability discovery in Linux-based firmware.
Introduces AgentFuzz, a directed greybox fuzzing framework for finding taint-style vulnerabilities in real-world LLM-based agents.
Figure 3 · PDF p.8
Studies nonsense-keyword spear scams in search engines and evaluates mitigation paths.
Figure 1 · PDF p.2Studies security risks and defenses for agents and foundation models.
Figure 2 · PDF p.4Measures security debt in LLM-agent applications by studying known vulnerabilities and the trade-offs introduced by mitigation strategies.
Studies security risks and defenses for agents and foundation models.
Measures malicious and abusive uses of serverless cloud functions in the wild.
Figure 2 · PDF p.3Finds cross-site scripting vulnerabilities with local path-persistent fuzzing.
Figure 3 · PDF p.7
Evaluates whether jailbreak defenses improve safety without degrading model utility, highlighting persistent trade-offs in practical LLM defense.
Figure 1 · PDF p.2Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Figure 4 · PDF p.6
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Figure 1 · PDF p.3Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Figure 1 · PDF p.2Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Figure 1 · PDF p.3Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Figure 1 · PDF p.5
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Figure 1 · PDF p.2Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Figure 1 · PDF p.6Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Figure 1 · PDF p.2Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Figure 2 · PDF p.7Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Figure 1 · PDF p.2
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Figure 1 · PDF p.3Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Figure 1 · PDF p.2
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Figure 1 · PDF p.2Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Figure 1 · PDF p.4Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Systematically measures privacy leakage risks in general-purpose language models.
Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Directly tests whether complete security patches are present in Java executables.
Figure 2 · PDF p.6Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Figure 1 · PDF p.3Studies robustness, backdoors, watermarks, poisoning, and trust in machine learning models.
Figure 1 · PDF p.3Studies security risks and defenses for agents and foundation models.
Figure 1 · PDF p.6Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Figure 1 · PDF p.1Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Figure 3 · PDF p.5Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Investigates privacy, abuse ecosystems, and real-world cybersecurity threats.
Advances vulnerability discovery, testing, and protection for software and intelligent systems.
Develops learning methods and AI systems across recommendation, vision, and time-series modeling.
Starting with limited target information, AgentCyberRange evaluates how agents discover attack surfaces, exploit vulnerabilities, and carry access, credentials, and network reachability through a multistage attack.
Controlled experiments examining agents' capacity for self-replication, their behavior, and the limits of human control.
NUWA essay on making frontier AI risk evaluation more scientific, evidence-based, and operational.
RESEARCH MISSION
NUWA Frontier AI Safety Lab was initiated and is supported by Whitzard. Its research covers frontier risk, agent and model safety, and risk evaluation and governance.
Study risks from increasingly capable and autonomous AI systems and develop verifiable approaches to control.
Study systemic safety questions in reasoning, action, tool use, model behavior, and human–AI trust.
Build executable, reproducible evaluations and study how evidence should inform risk assessment and governance decisions.
SELF-REPLICATION RESEARCH
RESEARCH INFRASTRUCTURE
Public evaluation environments and benchmarks that make agent cybersecurity capability observable.
Evaluates frontier AI systems' autonomous cyber-attack capabilities in complex enterprise environments.
Official site ↗Synthesizes executable test environments that combine deterministic code state with narrative dynamics for frontier AI risk evaluation.
View paper ↗RESEARCH & PUBLIC IMPACT
NUWA Lab is deeply involved in international AI safety consensus, research priorities, and generative AI safety standards.
CONSENSUS & RESEARCH PRIORITIES
Alignment · Human control · Verifiable red lines
Deep involvement from NUWA Lab
An international scientific consensus on alignment, human control, safety assurance, and verifiable behavioural red lines for advanced AI.
CONSENSUS & RESEARCH PRIORITIES
Risk assessment · Safe development · Control · Societal resilience
Deep involvement from NUWA Lab
A global agenda for urgent AI safety research shaped by contributors from frontier developers, government safety institutes, academia, and civil society.
NATIONAL STANDARD
Cybersecurity technology—Basic security requirements for generative artificial intelligence service
Deep involvement from NUWA Lab
A national baseline for training-data security, model security, service safeguards, and security assessment of generative AI services.
RECOGNITION
Top-tier cybersecurity conference
International interdisciplinary science breakthrough selection
Top-tier cybersecurity conference
Top-tier software-engineering conference
Top-tier cybersecurity conference
COLLABORATION
Collaborate on frontier-risk evaluation, agent safety, systems security, cybersecurity, and privacy.
Contact research
