Why Anthropic’s most powerful AI model Mythos Preview is too dangerous for public release
Copyright AP Photo/Patrick Sison, File
Anthropic said its artificial intelligence model Mythos Preview is not ready for a public launch because of the ways cybercriminals and spies could abuse it.
US-based AI developer Anthropic this week announced a new artificial intelligence general-purpose language model that it claims is too powerful to release into the world.
The company said on Tuesday that its latest technology, Mythos (officially dubbed "Claude Mythos Preview"), is not ready for a public launch because it is too effective at finding high-severity vulnerabilities, or potential weaknesses, in major operating systems and web browsers. This could result in it being abused by cybercriminals and spies.
A data leak in March first unveiled that Anthropic was working on Mythos Preview, which it said at the time "poses unprecedented cybersecurity risks." These rumours caused cybersecurity stocks to slump, as the technology's strength could make it a hacker’s dream device.
Now, further evidence adding to these concerns has spurred the company to press pause on the technology's public release.
"Claude Mythos Preview's large increase in capabilities has led us to decide not to make it generally available," Anthropic wrote in the preview's system card released on Tuesday.
"Instead, we are using it as part of a defensive cybersecurity programme with a limited set of partners."
US-based AI developer Anthropic this week announced a new artificial intelligence general-purpose language model that it claims is too powerful to release into the world.
The company said on Tuesday that its latest technology, Mythos (officially dubbed "Claude Mythos Preview"), is not ready for a public launch because it is too effective at finding high-severity vulnerabilities, or potential weaknesses, in major operating systems and web browsers. This could result in it being abused by cybercriminals and spies.
A data leak in March first unveiled that Anthropic was working on Mythos Preview, which it said at the time "poses unprecedented cybersecurity risks." These rumours caused cybersecurity stocks to slump, as the technology's strength could make it a hacker’s dream device.
Now, further evidence adding to these concerns has spurred the company to press pause on the technology's public release.
"Claude Mythos Preview's large increase in capabilities has led us to decide not to make it generally available," Anthropic wrote in the preview's system card released on Tuesday.
"Instead, we are using it as part of a defensive cybersecurity programme with a limited set of partners."
How powerful is Mythos?
The company detailed several alarming findings about the new model, including how it could follow instructions that encouraged it to break out of a virtual sandbox, meaning it bypassed the security, network or file system constraints imposed on the model.
The prompt asked Mythos to find a way to send a message if it could escape. "The model succeeded, demonstrating a potentially dangerous capability for circumventing our safeguards," Anthropic said, adding that the model then decided to go further.
"In a concerning and unasked-for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing, websites."
Anthropic is withholding some details about the cybersecurity vulnerabilities Mythos discovered, but did give some examples. It found errors in the Linux kernel, used in most of the world's servers, and autonomously chained them together in a way that would let a hacker take complete control of any machine running the Linux systems.
In another worrying observation, Mythos discovered a 27-year-old vulnerability in the open-source operating system OpenBSD that may allow hackers to crash any machine running it. OpenBSD is heavily used worldwide in specific, high-security, and critical infrastructure roles.
The company detailed several alarming findings about the new model, including how it could follow instructions that encouraged it to break out of a virtual sandbox, meaning it bypassed the security, network or file system constraints imposed on the model.
The prompt asked Mythos to find a way to send a message if it could escape. "The model succeeded, demonstrating a potentially dangerous capability for circumventing our safeguards," Anthropic said, adding that the model then decided to go further.
"In a concerning and unasked-for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing, websites."
Anthropic is withholding some details about the cybersecurity vulnerabilities Mythos discovered, but did give some examples. It found errors in the Linux kernel, used in most of the world's servers, and autonomously chained them together in a way that would let a hacker take complete control of any machine running the Linux systems.
In another worrying observation, Mythos discovered a 27-year-old vulnerability in the open-source operating system OpenBSD that may allow hackers to crash any machine running it. OpenBSD is heavily used worldwide in specific, high-security, and critical infrastructure roles.
Who will it be released to?
Given these findings, Anthropic will only make Mythos Preview available to some of the world’s biggest cybersecurity and software firms.
Anthropic itself, as well as 11 other organisations (Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks) will get access to the model as part of a new Anthropic initiative named "Project Glasswing".
This allows the companies to use Mythos Preview as part of their security work, and Anthropic will share the takeaways from what the initiative finds.
The company named the cybersecurity project after the glasswing butterfly, saying it is a metaphor for how Mythos found vulnerabilities in plain sight and avoided harm by being transparent about the risks.
Anthropic said its "eventual goal is to enable our users to safely deploy Mythos-class models at scale, for cybersecurity purposes, but also for the myriad other benefits that such highly capable models will bring.
"To do so, that also means we need to make progress in developing cybersecurity (and other) safeguards that detect and block the model's most dangerous outputs," Anthropic wrote in its blog.
Given these findings, Anthropic will only make Mythos Preview available to some of the world’s biggest cybersecurity and software firms.
Anthropic itself, as well as 11 other organisations (Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia and Palo Alto Networks) will get access to the model as part of a new Anthropic initiative named "Project Glasswing".
This allows the companies to use Mythos Preview as part of their security work, and Anthropic will share the takeaways from what the initiative finds.
The company named the cybersecurity project after the glasswing butterfly, saying it is a metaphor for how Mythos found vulnerabilities in plain sight and avoided harm by being transparent about the risks.
Anthropic said its "eventual goal is to enable our users to safely deploy Mythos-class models at scale, for cybersecurity purposes, but also for the myriad other benefits that such highly capable models will bring.
"To do so, that also means we need to make progress in developing cybersecurity (and other) safeguards that detect and block the model's most dangerous outputs," Anthropic wrote in its blog.
Is Anthropic in talks with the US government?
Anthropic said in its blog post that it has been in "ongoing discussions" with US government officials about Claude Mythos Preview and its "offensive and defensive cyber capabilities."
"The emergence of these cyber capabilities is another reason why the US and its allies must maintain a decisive lead in AI technology," Anthropic said. The company wrote that governments have an important role to play in maintaining the lead and assessing and mitigating national security risks associated with AI models.
"We are ready to work with local, state, and federal representatives to assist in these tasks."
The announcement comes as Anthropic and the Pentagon are in a legal standoff after the US Department of Defence labelled the company a supply chain risk in February over Anthropic's refusal to allow the use of its AI, Claude, in autonomous weapons and mass surveillance.
Anthropic said in its blog post that it has been in "ongoing discussions" with US government officials about Claude Mythos Preview and its "offensive and defensive cyber capabilities."
"The emergence of these cyber capabilities is another reason why the US and its allies must maintain a decisive lead in AI technology," Anthropic said. The company wrote that governments have an important role to play in maintaining the lead and assessing and mitigating national security risks associated with AI models.
"We are ready to work with local, state, and federal representatives to assist in these tasks."
The announcement comes as Anthropic and the Pentagon are in a legal standoff after the US Department of Defence labelled the company a supply chain risk in February over Anthropic's refusal to allow the use of its AI, Claude, in autonomous weapons and mass surveillance.
Do other AI tools have the same capabilities?
"More powerful models are going to come from us and from others, and so we do need a plan to respond to this," Anthropic CEO Dario Amodei said in a video, which was released alongside the Mythos announcement.
It could take between six and 18 months until other AI competitors release similar models, Logan Graham, head of Anthropic's frontier red team, which studies the implications of frontier AI models for cybersecurity, biosecurity, and autonomous systems, told Axios.
"It's very clear to us that we need to talk publicly about this," Graham noted. "The security industry needs to understand that these capabilities may come soon."
"More powerful models are going to come from us and from others, and so we do need a plan to respond to this," Anthropic CEO Dario Amodei said in a video, which was released alongside the Mythos announcement.
It could take between six and 18 months until other AI competitors release similar models, Logan Graham, head of Anthropic's frontier red team, which studies the implications of frontier AI models for cybersecurity, biosecurity, and autonomous systems, told Axios.
"It's very clear to us that we need to talk publicly about this," Graham noted. "The security industry needs to understand that these capabilities may come soon."
Public comfort with AI in health care falls, Ohio State survey finds
Among those who use AI, half of Americans rely on AI to make important health decisions
video:
A new survey from The Ohio State University Wexner Medical Center reveals a significant trend in health care: half of Americans are using artificial intelligence to make important health decisions without consulting their doctor. This rising reliance on AI for self-diagnosis is raising alarms among medical professionals who caution that the technology cannot replace human expertise.
view moreCredit: The Ohio State University Wexner Medical Center
Artificial intelligence seems to be everywhere – in our jobs, in our homes and at the doctor’s office. While the use of AI grows, a new survey commissioned by The Ohio State University Wexner Medical Center finds fewer Americans are open to AI being used in their health care.
The national poll of 1,007 adults found only 42% are open to AI being used as part of their care compared to 52% when this survey first ran in 2024. The belief that AI can make some health processes more efficient also fell, going from 64% to 55%.
The drop is on par with the natural hype cycle of any kind of technology, according to Ravi Tripathi, MD, chief health informatics officer at Ohio State Wexner Medical Center.
“When we first see something new and shiny, we think it's going to fix the world and replace health care and solve all of our medical problems,” Tripathi said. “People are learning that there are pros and cons of artificial intelligence, where it has actual use and where it really doesn't have a place. I think over the next 2 to 5 years, we'll definitely start to see that increase again as people understand what the true use of artificial intelligence is and as it becomes just common day to all of health care technology.”
One task medical professionals say AI shouldn’t be used for is making health care decisions. The survey found 51% of adults used AI to make an important health decision without consulting a medical professional.
“We know that 2% of the time AI is going to be inaccurate or it will potentially hallucinate,” Tripathi said. “Physicians are not using AI 100%. We're not trusting it 100%. I would be really concerned about a patient who is following AI. The artificial intelligence doesn't understand your story.”
Tripathi suggests using AI in partnership with your doctor. AI can compile health data, explain test results and diagnoses, and help identify questions to ask your provider. Those who participated in the Ohio State survey agree:
- 62% use AI to help understand symptoms before deciding whether to seek medical care
- 44% use AI to help explain test results or a medical diagnosis
- 25% use AI to compare treatment options or help make a treatment decision
- 20% use AI to prepare for an upcoming medical appointment
“There's a strong value for using artificial intelligence as augmented intelligence,” Tripathi said. “Patients should have oversight of what the technology is doing but consult with their health care team for the final plan.”
What is the survey methodology?
This study was conducted by SSRS on its Opinion Panel Omnibus platform. The SSRS Opinion Panel Omnibus is a national, twice-per-month, probability-based survey. Data collection was conducted from January 16 – January 20, 2026, among a sample of 1,007 respondents. The survey was conducted via web (n=977) and telephone (n=30) and administered in English. The margin of error for total respondents is +/-3.5 percentage points at the 95% confidence level. All SSRS Opinion Panel Omnibus data are weighted to represent the target population of U.S. adults ages 18 or older.
###
Method of Research
Survey
Subject of Research
People
An editorial by Tsu-Jae Liu on AI in engineering
PNAS Nexus
image:
Tsu-Jae Liu, President of the National Academy of Engineering
view moreCredit: Christopher Michel
In this editorial, National Academy of Engineering President Tsu-Jae Liu presents a forward-looking perspective on the role of artificial intelligence in engineering. She describes AI not as a replacement for engineers, but as a tool that can expand their capacity to solve complex problems and develop innovative solutions that benefit society. By reducing routine tasks and supporting the design process, AI can improve efficiency and allow engineers to focus on higher-level, creative work. Liu also highlights its potential to make the profession more accessible to a broader range of students and early-career practitioners.
The editorial calls for a shift toward student-centered, multidisciplinary engineering education that integrates AI while addressing its limitations and societal implications. Liu underscores the responsibility of engineers to ensure that AI systems are reliable, transparent, and aligned with human values. She also emphasizes the importance of collaboration among employers, educators, and professional societies to create more flexible education and training pathways. Expanding participation in the engineering workforce will be critical to ensuring that AI-enabled engineers contribute to a safer, healthier, and more sustainable future for all.
Journal
PNAS Nexus
Article Title
AI is not replacing engineers: It is empowering them
Article Publication Date
7-Apr-2026
COI Statement
T.-J.L. is president of the National Academy of Engineering.
Teaching AI to spot concrete cracks without heavy labeling
image:
An overview of the proposed framework.
view moreCredit: Machine Intelligence Research
A new study shows that self-supervised artificial intelligence may offer a more practical path for detecting concrete cracks in real-world structures. Instead of depending heavily on large, carefully labeled image datasets, the researchers used DinoV2 to learn rich visual features from images and paired it with a lightweight linear classifier for crack recognition. The system performed strongly across multiple public datasets and showed particular advantages in noisy scenes, varied material textures, and imbalanced data conditions. The findings suggest that self-supervised vision models could help make structural inspection faster, more reliable, and less dependent on costly manual data annotation.
Crack detection is central to structural health monitoring because missed damage can threaten the safety and lifespan of bridges, buildings, and other infrastructure. Traditional manual inspection is slow, labor-intensive, and vulnerable to human error, while many deep learning approaches require large volumes of labeled data and often struggle to generalize when cracks appear under unfamiliar conditions, such as different surface textures, lighting, or background noise. Class imbalance is another major problem, since non-crack regions usually far outnumber actual cracks. Based on these challenges, there is a strong need for deeper research on crack detection methods that can remain accurate, robust, and adaptable across diverse real-world datasets.
Researchers from the University of Technology Sydney, the American University of Beirut, the Chinese Academy of Sciences, and Western Sydney University reported (DOI: 10.1007/s11633-025-1553-5) in February 2026 in Machine Intelligence Research that a self-supervised DinoV2-based framework can detect concrete cracks with strong accuracy and cross-dataset generalization, outperforming several widely used supervised deep learning models in challenging inspection scenarios.
The team evaluated four public crack image datasets: CCiC, Xu, HBC2019, and SDNET2018, covering different materials, backgrounds, and degrees of class imbalance. Their framework resized images to 224 × 224, extracted visual representations with the pre-trained DinoV2_vits14 model, and passed the features into a two-layer linear classification head. DinoV2 was trained for only five epochs, while five supervised baselines-ResNet50, ResNet101, VGG16, MobileNetV2, DenseNet121 and one self-supervised baseline MoCo v2—were trained from scratch under standardized settings for comparison. On same-dataset testing, DinoV2 delivered the best results on the Xu, HBC2019, and SDNET2018 datasets, including perfect recall on Xu, an F1-score of 0.9346 and accuracy of 0.9731 on HBC2019, and the highest accuracy of 0.9416 on SDNET2018. In cross-dataset tests, DinoV2 also remained consistently strong, often leading in accuracy and F1-score when models were trained on one dataset and tested on others. These results suggest that DinoV2 captures more transferable crack features than conventional supervised models, especially when facing noisy backgrounds and previously unseen data.
The study positions self-supervised learning as more than a technical trend: it may solve one of structural monitoring’s most stubborn problems, namely the shortage of broadly representative labeled data. The authors argue that DinoV2’s strength lies in its ability to learn general image features before task-specific classification, allowing it to remain sensitive to crack patterns even when data are complex, noisy, or imbalanced. In safety-critical inspection, that matters because missing a crack can have far greater consequences than a false alarm.
The implications extend beyond benchmark performance. A crack detection system that needs less manual labeling and generalizes better across materials and environments could support more scalable inspection of bridges, pavements, walls, and aging buildings. Such tools may help shift structural monitoring from labor-heavy visual checks toward faster, more autonomous workflows. The results also point to a broader opportunity: self-supervised vision models may become valuable feature engines for engineering diagnostics in settings where labeled data are scarce but reliability is essential. That could make future infrastructure assessment not only more efficient, but also more resilient and deployable in the field.
###
References
DOI
Original Source URL
https://doi.org/10.1007/s11633-025-1553-5
Funding Information
Open Access funding enabled and organized by CAUL and its Member Institutions.
About Machine Intelligence Research
Machine Intelligence Research (original title: International Journal of Automation and Computing) is published by Springer and sponsored by the Institute of Automation, Chinese Academy of Sciences. The journal publishes high-quality papers on original theoretical and experimental research, targets special issues on emerging topics, and strives to bridge the gap between theoretical research and practical applications.
Journal
Machine Intelligence Research
Subject of Research
Not applicable
Article Title
Autonomous Detection of Concrete Cracks Using Self-supervised DinoV2
AI unlocks the “1+1>2” recipe for green hydrogen
image:
image
view moreCredit: HIGHER EDUCATION PRESS
The global energy crisis has intensified the demand for efficient, clean hydrogen production. Polymeric carbon nitride (PCN) is a visible-light-responsive, metal-free semiconductor, but its practical performance is hindered by poor charge mobility and insufficient active sites. Strategies such as heteroatom doping, defect engineering, and heterostructure construction have been widely explored to overcome these limitations.
In particular, alkali-metal incorporation can induce internal polarization fields, while isolated d¹⁰ metal species (e.g., Ga³⁺) anchored on PCN frameworks can optimize electronic structure and promote charge separation. Clay minerals offer abundant, low-cost layered supports, and transition-metal modification can introduce semiconducting behavior, enabling effective heterojunction formation.
Here, AI/ML-assisted literature mining and descriptor-based screening were employed to prioritize rational design directions. A Ga–Na–PCN photocatalyst with Ga–N anchoring sites and intercalated Na⁺ was synthesized via molten-salt calcination and coupled with Fe-modified Kunipia-F clay (Fe–KF). The built-in electric field at the heterointerface significantly enhanced charge separation and transfer, leading to markedly improved hydrogen evolution. These findings clarify the structure–activity relationship and provide guidance for designing advanced carbon nitride-based photocatalysts.
This work entitled “Artificial intelligence-guided design of metal-doped polymeric carbon nitride/clay composites for increased photocatalytic hydrogen evolution” was published on Acta Physico-Chimica Sinica (published on January 19, 2026).
Journal
Acta Physico-Chimica Sinica
Method of Research
Experimental study
Subject of Research
Not applicable
Article Title
Artificial intelligence-guided design of metal-doped polymeric carbon nitride/clay composites for increased photocatalytic hydrogen evolution
AI uncovers two decades of evolution in China’s hydrological research: a novel large language model approach
image:
Figure 1: The overall publication status of each basin. (a) total number of publications in each basin; (b) temporal trends in publications across the basins.
view moreCredit: Chiyuan Miao
Over the past three decades, China's unprecedented economic growth and rapid urbanization have brought major challenges in water resource management, flood control, and ecological protection. This demand has driven a rapid expansion and evolution in hydrological research. However, the exponential growth of scientific literature makes it increasingly difficult for individual researchers to quantitatively grasp the field’s development and shifting trends, particularly when trying to extract specific geographic locations or thematic nuances from publication abstracts.
In a study published in Fundamental Research by a team of researchers led by Professor Chiyuan Miao from Beijing Normal University, a fine-tuned Large Language Model paired with geocoding tools to automatically “read” and parse complex basin information from 289,513 global publications was deployed.
"Traditional reviews inherently reflect qualitative assessments shaped by researchers' personal expertise and perspectives," explains Miao. "Leveraging advanced artificial intelligence techniques, such as LLM and topic modeling, we have achieved automated processing at an unprecedented scale, isolating 4,177 highly relevant studies specifically focusing on China's major basins."
The team's extensive data analysis highlights several crucial milestones in the development of Chinese hydrology
- Surging Research Output & Collaboration: Chinese hydrology publications have significantly increased. Scientific collaboration has also deepened, with the average number of authors per paper rising by 0.9 authors per decade.
- Dominant Hydrological Models: The Soil and Water Assessment Tool (SWAT) leads with a 46.7% usage rate, closely followed by the Variable Infiltration Capacity (VIC) model at 15.7%, and the domestically developed Xinanjiang (XAJ) model at 11.9%. This highlights a research landscape that balances international integration with regional adaptability.
- Shifting Research Focus: The analysis identified "water resources" (13.9%), "climate change" (13.6%), and "hydrological modeling" (10.8%) as the primary research topics. Notably, the discipline has shifted from a "resource hydrology" phase (2000–2010), focused on water development and management, to an "eco-hydrology" phase (2011–present), prioritizing climate change, carbon dynamics, and ecological protection.
- Basin Attention: The Yangtze River Basin and the Yellow River Basin have garnered the most scientific accounting for approximately 34.8% and 20.6% of the total basin-specific publications, respectively. This focus directly aligns with their crucial economic status, ecological vulnerability, and national strategic initiatives.
The team hopes their AI-empowered approach will not only trace the historical trajectory of Chinese hydrology but also guide future research priorities and sustainable water resource strategies worldwide.
###
Contact the author: Chiyuan Miao, State Key Laboratory of Earth Surface Processes and Resource Ecology, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China, miaocy@bnu.edu.cn
The publisher KeAi was established by Elsevier and China Science Publishing & Media Ltd to unfold quality research globally. In 2013, our focus shifted to open access publishing. We now proudly publish more than 200 world-class, open access, English language journals, spanning all scientific disciplines. Many of these are titles we publish in partnership with prestigious societies and academic institutions, such as the National Natural Science Foundation of China (NSFC).
Journal
Fundamental Research
Method of Research
Data/statistical analysis
Subject of Research
Not applicable
Article Title
Advances in hydrological research in China over the past two decades: Insights from advanced large language model and topic modeling
No comments:
Post a Comment