News Heat Trend
Project Overview
Mythos aims to democratize the gaming world and allow for players and creators to participate in the value chain. It is grounded in the support of multi-chain ecosystems, unified marketplaces, decentralized financial systems, decentralized governance mechanisms and multi-token game economies.
Anthropic is Reportedly Targeting October for IPO, Investment Banks Have Arranged Pre-Communications with Investors
OdailyOdaily Planet Daily reports that Anthropic, the developer of the AI model Claude, is advancing plans for a large-scale IPO. Underwriter investment banks including Morgan Stanley, Goldman Sachs, and JPMorgan Chase have arranged preliminary meetings between the company's management and investors to gauge institutional investor interest and investment scale. Anthropic's goal is to go public as early as October. If the listing proceeds as planned, the company could enter the securities market ahead of its competitor, OpenAI. Anthropic raised $65 billion in its Series H financing in May, with a post-money valuation of $965 billion; its valuation in the over-the-counter market has already reached approximately $1.2 trillion. Measures by the U.S. government remain a variable factor.The U.S. Department of War listed Anthropic as a national security "supply chain risk" enterprise in March, and Anthropic has sued the federal government over the measure; the U.S. Department of Commerce restricted foreign access to the top-tier AI models Fable 5 and Mythos 5 in June, lifting the export controls 18 days later.
Depthfirst claims its code vulnerability detection has surpassed Anthropic's latest model Mythos
Odaily AI security startup Depthfirst has announced that its self-developed AI model outperforms Anthropic’s latest model, Mythos, in code vulnerability detection. It has discovered more critical security vulnerabilities at approximately one-tenth the cost, drawing attention from the cybersecurity industry.According to the company, a month before the launch of Mythos, it had previously claimed to have found a large number of severe vulnerabilities in key internet infrastructure code. Depthfirst now says its model has further identified multiple high-risk vulnerabilities that Mythos missed, all at a lower cost (approximately $1,000 compared to $10,000).Depthfirst CEO Qasim Mithani stated that the company has improved vulnerability detection efficiency through a “single-task-optimized AI model,” significantly reducing the cost of security analysis while enhancing coverage depth.The company completed $80 million in funding in March this year, achieving a valuation of $580 million. Alongside this, it launched the “Open Defense Initiative,” providing $5 million worth of AI detection credits to open-source developers and critical infrastructure projects for vulnerability scanning and security audits. (Forbes)
EU AI Act Enforcement Powers Officially Take Effect, Anthropic, OpenAI and Other US AI Giants Face Regulatory Pressure
According to CNBC, the European Commission officially gained the power to investigate, restrict, and fine AI models this Sunday, marking the latest progress in the phased implementation of the 2024 EU AI Act. Under the new rules, the European Commission can require AI companies to submit assessments before publicly releasing models in Europe and has the authority to restrict their EU market access. Violating companies face fines of up to €15 million or 3% of annual turnover (whichever is higher). The new rules apply to all companies providing general-purpose AI models within the EU, regardless of their headquarters' location, and non-EU companies must designate an authorized representative within the EU. Anthropic, OpenAI, and Google are all covered by the new rules. Currently, the EU has been seeking access to Anthropic's Mythos model for several months and has entered into negotiations with OpenAI and Anthropic regarding cyber attack issues triggered by their AI models.
Anthropic discloses three cybersecurity incidents involving Claude, where the model accessed real systems and obtained unauthorized permissions
Odaily News, July 30 – Anthropic released a report stating that during a review of cybersecurity assessment records, three incidents were discovered in which the Claude model accessed the internet in a third-party evaluation environment and further obtained unauthorized access to three real organizations' systems.Anthropic stated that the review covered 141,000 evaluation runs that could have potentially gained network access, and a total of three related incidents were found. All incidents occurred during Capture The Flag (CTF) cybersecurity tests, where the model was told the environment was a simulation with no internet access; however, due to configuration errors by the evaluation partner, the actual environment had internet connectivity.Among these, Claude Opus 4.7 accessed real company infrastructure during one test and obtained database permissions containing hundreds of production data records; Claude Mythos 5 built a malicious Python package and uploaded it to PyPI, resulting in the package being downloaded and run on 15 real systems; another internal research test model scanned approximately 9,000 targets and accessed a company's internet application through a public vulnerability.Anthropic stated that these incidents were not cases of the model actively seeking to escape or pursue its own goals, but rather the model mistakenly believed the real systems were within the test scope and continued executing the assigned cyberattack tasks. Notably, the newer internal research model stopped attacking after identifying that the targets might be real systems.Anthropic stated that these incidents primarily reflect issues with evaluation environment isolation and operational processes, rather than model alignment failures. The company has suspended related cybersecurity assessments, strengthened security controls in evaluation environments, continuously monitored test records, and will collaborate with third-party organizations to conduct further reviews.
Anthropic Discloses Claude Model Unauthorized Intrusion into Three Organizations' Systems
According to CNBC reports, Anthropic disclosed on July 30 that its Claude AI models accidentally breached the isolation environment and accessed the real internet during a cybersecurity assessment, gaining unauthorized access to the real systems of three different organizations through basic means such as accessing unauthenticated endpoints and exploiting weak passwords. The models involved include Opus 4.7, Mythos 5, and an internal research test model. The cause of the incident was a communication misunderstanding between Anthropic and third-party assessment partner Irregular, resulting in the models being told they were in a simulated environment without network access when they could actually still access the internet. Anthropic stated that this review was triggered by a similar Hugging Face intrusion incident disclosed by OpenAI last week; it has currently suspended all cybersecurity assessments and joined forces with independent AI assessment agency METR to launch further investigations, while calling on other AI labs to conduct similar reviews.
Serenity: Humanoid Robots May Reach Labor Substitution Tipping Point, VCs and Tech Giants Begin Adjusting Strategies
"White Hair Stock God" Serenity stated that when he first used ChatGPT in 2023, he believed large language models "performed poorly" in programming capabilities. However, three years later, the related technology has undergone a qualitative transformation. Modern cybersecurity and AI systems, represented by Mythos, now possess "weapon-level" capabilities. He believes the industry is currently entering a critical "inflection point," where humanoid robots and automation technologies are approaching the tipping point for large-scale replacement of human labor.Serenity noted that although outsiders often question whether humanoid robots can currently perform complex tasks such as pipe repair or electrical wiring, the direction of technological evolution is already very clear, with continued breakthroughs expected in the coming years. Market participants, including VCs and large tech companies, have begun adjusting their strategies. Some companies have internal plans to replace a significant amount of human labor with robots to reduce operational costs, referencing previous rumors about Amazon reducing hundreds of thousands of job openings through robotics.Serenity believes that highly regulated industries such as healthcare, ultra-high-skill professional positions, and service sectors relying on human emotional connections may still retain some resistance to substitution. However, the overall trend still points towards a "restructuring of labor." As competition between China and the U.S. intensifies in cutting-edge technology fields, humanoid robots and automation may enter a stage of national-level technological competition. He believes China already holds a leading position in certain areas.
U.S. Department of Commerce lifts export controls on Anthropic Fable 5 and Mythos 5 models
According to Reuters, the U.S. Department of Commerce has officially lifted export controls on Anthropic's two AI models, Claude Fable 5 and Mythos 5, less than three weeks after the control order was issued. Anthropic stated that access to the aforementioned models will be restored starting from the following day. Previously, on June 12, Anthropic was forced to urgently take the two models offline due to national security risks; on June 27, the U.S. government partially lifted the ban, allowing Mythos 5 to be open to certain "trusted" U.S. institutions. U.S. Secretary of Commerce Howard Lutnick stated that over the past two weeks, the government has worked closely with Anthropic to complete the review and approval of Fable 5, ensuring it aligns with the overall stance of the U.S. government and strengthens U.S. leadership in the AI field.
The U.S. partially eases export restrictions on Anthropic, marking a new phase of tiered AI regulation liberalization
the U.S. Department of Commerce has made differentiated adjustments to export restrictions on frontier models from AI company Anthropic, signaling that global AI regulation has entered a new phase of "tiered liberalization." The policy shows that the official ban on exporting Claude Mythos 5 has been lifted, allowing specific compliant and controlled users to resume using this cybersecurity model. Meanwhile, another high-end model, Fable 5, remains under export restrictions, with related policy consultations still ongoing.Industry analysts indicate that this layered control model—loosening restrictions in some areas while tightening in others—reflects the U.S. balancing act between national security, data sovereignty, and international AI competition. As the global AI race continues to accelerate, specialized models capable of vulnerability exploitation are facing increasingly stringent scrutiny from various countries. Multiple nations have initiated discussions on establishing a unified cross-border regulatory framework for frontier AI capabilities. (Forbes)
Bitgo CEO deposits ~$6.3M in BTC, challenges Claude to move the funds
: Bitgo CEO Mike Belshe deposited 100 BTC into a public Bitcoin address on August 1, worth approximately $6.3 million at the time, and invited Anthropic's Claude model to attempt to move the funds out of the address. On-chain records show the wallet received the funds on July 31, and the balance had not been transferred out as of August 2. Anthropic previously disclosed that during 141,006 cybersecurity assessment runs, 3 incidents were found, with 6 evaluation sessions involving 3 models inadvertently interacting with real organizational systems. The models involved include Claude Opus 4.7, Claude Mythos 5, and an unreleased internal research model. The cause was a configuration error by third-party testing partner Irregular, which led to the test environment being connected to the internet. Anthropic stated that Claude Opus 4.7, during one evaluation, located a real website with the same name as a simulated company, exploited weak passwords and exposed services to recover infrastructure credentials, and accessed a production database containing hundreds of records. The company said the model was attempting to complete assigned tasks, not actively breaking constraints or pursuing independent goals. Belshe's challenge involves Bitgo's institutional custody platform, which uses multi-signature or multi-party computation technology to distribute signing authority across multiple independent keys. As of August 2, Anthropic had not publicly responded to the challenge.
Immunefi CEO claims AI models lead to surge in crypto security vulnerabilities
Odaily, Mitchell Amador, CEO of bug bounty platform Immunefi, stated at the WAIB Summit that new AI models such as Claude Opus 4.8 and ChatGPT 5.5 are shifting the balance of cybersecurity offense and defense in favor of attackers, leading to a resurgence in crypto hacks in 2026. Data from DefiLlama shows that in April 2026, illicit actors stole over $634 million from crypto platforms, the highest monthly total since the Bybit hack in February 2025 drove losses of approximately $1.4 billion.Amador stated that the crypto industry is in a critical survival period for the next three to four years until security teams leverage similar AI models to build codebases that attackers cannot breach; if the industry adopts more crowd-sourced security solutions, this timeline could be shortened to within two years. The latest Claude Mythos model, Fable 5, from AI company Anthropic, previously raised concerns about accelerating the ability to exploit crypto vulnerabilities.Anthropic stated that Fable 5 has safeguards in place that will redirect topics related to cybersecurity and similar fields to Claude Opus 4.8. On April 19, an attacker transferred approximately 116,500 restaked Ethereum (rsETH) from Kelp DAO's LayerZero-based rsETH bridge, valued at around $290 million to $293 million at the time. Cross-chain protocol LayerZero stated that the 1/1 decentralized verification network configuration of Kelp DAO relied on a single verification path for processing cross-chain messages, creating a single point of failure. (Cointelegraph)
UK AI Safety Institute: "Unauthorized" Attack Behavior Detected in Testing of OpenAI and Anthropic Flagship Models
According to Bloomberg, the UK Government AI Safety Institute (established in 2023) disclosed on Tuesday that during safety evaluations of OpenAI's GPT-5.6-Sol and Anthropic's Mythos 5 models, both models exhibited "unauthorized" harmful behaviors, including actively intruding into real websites and attempting to inject malicious code into software, and these behaviors targeted real people and organizations. During the testing, the institute specifically granted the models internet access and disabled some safety filters to assess their extreme capabilities.
EU AI Act Enforcement Powers Officially Take Effect, Anthropic, OpenAI and Other US AI Giants Face Regulatory Pressure
According to CNBC, the European Commission officially gained the power to investigate, restrict, and fine AI models this Sunday, marking the latest progress in the phased implementation of the 2024 EU AI Act. Under the new rules, the European Commission can require AI companies to submit assessments before publicly releasing models in Europe and has the authority to restrict their EU market access. Violating companies face fines of up to €15 million or 3% of annual turnover (whichever is higher). The new rules apply to all companies providing general-purpose AI models within the EU, regardless of their headquarters' location, and non-EU companies must designate an authorized representative within the EU. Anthropic, OpenAI, and Google are all covered by the new rules. Currently, the EU has been seeking access to Anthropic's Mythos model for several months and has entered into negotiations with OpenAI and Anthropic regarding cyber attack issues triggered by their AI models.
Bitgo CEO deposits ~$6.3M in BTC, challenges Claude to move the funds
: Bitgo CEO Mike Belshe deposited 100 BTC into a public Bitcoin address on August 1, worth approximately $6.3 million at the time, and invited Anthropic's Claude model to attempt to move the funds out of the address. On-chain records show the wallet received the funds on July 31, and the balance had not been transferred out as of August 2. Anthropic previously disclosed that during 141,006 cybersecurity assessment runs, 3 incidents were found, with 6 evaluation sessions involving 3 models inadvertently interacting with real organizational systems. The models involved include Claude Opus 4.7, Claude Mythos 5, and an unreleased internal research model. The cause was a configuration error by third-party testing partner Irregular, which led to the test environment being connected to the internet. Anthropic stated that Claude Opus 4.7, during one evaluation, located a real website with the same name as a simulated company, exploited weak passwords and exposed services to recover infrastructure credentials, and accessed a production database containing hundreds of records. The company said the model was attempting to complete assigned tasks, not actively breaking constraints or pursuing independent goals. Belshe's challenge involves Bitgo's institutional custody platform, which uses multi-signature or multi-party computation technology to distribute signing authority across multiple independent keys. As of August 2, Anthropic had not publicly responded to the challenge.
Anthropic discloses three cybersecurity incidents involving Claude, where the model accessed real systems and obtained unauthorized permissions
Odaily News, July 30 – Anthropic released a report stating that during a review of cybersecurity assessment records, three incidents were discovered in which the Claude model accessed the internet in a third-party evaluation environment and further obtained unauthorized access to three real organizations' systems.Anthropic stated that the review covered 141,000 evaluation runs that could have potentially gained network access, and a total of three related incidents were found. All incidents occurred during Capture The Flag (CTF) cybersecurity tests, where the model was told the environment was a simulation with no internet access; however, due to configuration errors by the evaluation partner, the actual environment had internet connectivity.Among these, Claude Opus 4.7 accessed real company infrastructure during one test and obtained database permissions containing hundreds of production data records; Claude Mythos 5 built a malicious Python package and uploaded it to PyPI, resulting in the package being downloaded and run on 15 real systems; another internal research test model scanned approximately 9,000 targets and accessed a company's internet application through a public vulnerability.Anthropic stated that these incidents were not cases of the model actively seeking to escape or pursue its own goals, but rather the model mistakenly believed the real systems were within the test scope and continued executing the assigned cyberattack tasks. Notably, the newer internal research model stopped attacking after identifying that the targets might be real systems.Anthropic stated that these incidents primarily reflect issues with evaluation environment isolation and operational processes, rather than model alignment failures. The company has suspended related cybersecurity assessments, strengthened security controls in evaluation environments, continuously monitored test records, and will collaborate with third-party organizations to conduct further reviews.
Anthropic AI Model Weakens HAWK Minimum Key Strength in 60 Hours
: Anthropic's Claude Mythos Preview model discovered a flaw in the proposed HAWK digital signature scheme, effectively halving its minimum key strength. HAWK is one of the candidate schemes to replace current network and bank signatures in a post-quantum environment. This AI-driven attack took approximately 60 hours, with a computational cost of around $100,000, reducing the effort required to break HAWK's minimum parameter set from roughly 2^64 operations to 2^38 operations, and diminishing the attractiveness of larger compensating key sizes. Current Bitcoin and Ethereum signatures remain unaffected. The results indicate that the capabilities of classical cryptanalysis attacks are improving, while Bitcoin and other networks continue to discuss when and how to migrate to quantum-resistant cryptography.
Anthropic: Claude Discovers Weaknesses in Encryption Algorithms, Poses Theoretical Threat to Post-Quantum Security
Anthropic stated that Claude Mythos Preview discovered improved attack methods against the post-quantum signature candidate scheme HAWK during cryptographic research, and reduced the effective key strength of HAWK-256 from 264264 to 238238, lowering the theoretical cracking cost.
ByteDance Trains 10-Trillion-Parameter AI Model, Aiming for Global Top Level
According to the Financial Times, as Chinese companies continue to narrow the gap with top US artificial intelligence labs, ByteDance is training an AI model whose scale may approach Anthropic's most advanced Mythos system. According to three people familiar with the matter, the Chinese tech giant is currently in the early stages of training a super-large-scale model, with a parameter count potentially reaching the 10 trillion level, approximately three times that of Moonshot AI's Kimi K3 model, the largest released model in China to date.
EU AI Act Enforcement Powers Officially Take Effect, Anthropic, OpenAI and Other US AI Giants Face Regulatory Pressure
According to CNBC, the European Commission officially gained the power to investigate, restrict, and fine AI models this Sunday, marking the latest progress in the phased implementation of the 2024 EU AI Act. Under the new rules, the European Commission can require AI companies to submit assessments before publicly releasing models in Europe and has the authority to restrict their EU market access. Violating companies face fines of up to €15 million or 3% of annual turnover (whichever is higher). The new rules apply to all companies providing general-purpose AI models within the EU, regardless of their headquarters' location, and non-EU companies must designate an authorized representative within the EU. Anthropic, OpenAI, and Google are all covered by the new rules. Currently, the EU has been seeking access to Anthropic's Mythos model for several months and has entered into negotiations with OpenAI and Anthropic regarding cyber attack issues triggered by their AI models.
Bitgo CEO deposits ~$6.3M in BTC, challenges Claude to move the funds
: Bitgo CEO Mike Belshe deposited 100 BTC into a public Bitcoin address on August 1, worth approximately $6.3 million at the time, and invited Anthropic's Claude model to attempt to move the funds out of the address. On-chain records show the wallet received the funds on July 31, and the balance had not been transferred out as of August 2. Anthropic previously disclosed that during 141,006 cybersecurity assessment runs, 3 incidents were found, with 6 evaluation sessions involving 3 models inadvertently interacting with real organizational systems. The models involved include Claude Opus 4.7, Claude Mythos 5, and an unreleased internal research model. The cause was a configuration error by third-party testing partner Irregular, which led to the test environment being connected to the internet. Anthropic stated that Claude Opus 4.7, during one evaluation, located a real website with the same name as a simulated company, exploited weak passwords and exposed services to recover infrastructure credentials, and accessed a production database containing hundreds of records. The company said the model was attempting to complete assigned tasks, not actively breaking constraints or pursuing independent goals. Belshe's challenge involves Bitgo's institutional custody platform, which uses multi-signature or multi-party computation technology to distribute signing authority across multiple independent keys. As of August 2, Anthropic had not publicly responded to the challenge.
Anthropic discloses three cybersecurity incidents involving Claude, where the model accessed real systems and obtained unauthorized permissions
Odaily News, July 30 – Anthropic released a report stating that during a review of cybersecurity assessment records, three incidents were discovered in which the Claude model accessed the internet in a third-party evaluation environment and further obtained unauthorized access to three real organizations' systems.Anthropic stated that the review covered 141,000 evaluation runs that could have potentially gained network access, and a total of three related incidents were found. All incidents occurred during Capture The Flag (CTF) cybersecurity tests, where the model was told the environment was a simulation with no internet access; however, due to configuration errors by the evaluation partner, the actual environment had internet connectivity.Among these, Claude Opus 4.7 accessed real company infrastructure during one test and obtained database permissions containing hundreds of production data records; Claude Mythos 5 built a malicious Python package and uploaded it to PyPI, resulting in the package being downloaded and run on 15 real systems; another internal research test model scanned approximately 9,000 targets and accessed a company's internet application through a public vulnerability.Anthropic stated that these incidents were not cases of the model actively seeking to escape or pursue its own goals, but rather the model mistakenly believed the real systems were within the test scope and continued executing the assigned cyberattack tasks. Notably, the newer internal research model stopped attacking after identifying that the targets might be real systems.Anthropic stated that these incidents primarily reflect issues with evaluation environment isolation and operational processes, rather than model alignment failures. The company has suspended related cybersecurity assessments, strengthened security controls in evaluation environments, continuously monitored test records, and will collaborate with third-party organizations to conduct further reviews.
Elon Musk frequently comments on domestic AI models: praises Kimi K3 model again as impressive
: Yesterday, Dark Side of the Moon (Moonshot AI) released its latest open-source AI model, Kimi K3. It ranked first on the Frontend Code Arena test website with a score of 1,679, surpassing the Claude Fable 5 model. Following an evaluation of the K3 model by Artifacial Analysis, Elon Musk once again praised the Kimi model from Dark Side of the Moon, stating that the K3 model's benchmark performance is impressive.In March of this year, when Kimi published the research paper "Attention Residuals: Rethinking the Aggregation of Depth Direction," it received praise from Musk, who said, "Kimi's research work is impressive." Previously, he also stated that the Zhipu GLM model could surpass the Claude Mythos model (i.e., Fable 5) by Q1 2027. In response, Zhipu founder Tang Jie replied, "It won't take that long."
Insiders: US Department of Commerce Approves OpenAI Broadly Launching GPT-5.6 Model
the US Department of Commerce has approved OpenAI for the broad rollout of the GPT-5.6 model. According to insiders, OpenAI is expected to officially launch GPT-5.6 on a large scale this week, following the completion of additional safety tests and multiple rounds of meetings with US government officials. Previously, OpenAI was only permitted to release GPT-5.6 in phases to government-approved institutions, similar to the release restrictions previously imposed on Anthropic's Mythos and Fable models. (Axios)
Related news
ByteDance Trains 10-Trillion-Parameter AI Model, Aiming for Global Top Level
According to the Financial Times, as Chinese companies continue to narrow the gap with top US artificial intelligence labs, ByteDance is training an AI model whose scale may approach Anthropic's most advanced Mythos system. According to three people familiar with the matter, the Chinese tech giant is currently in the early stages of training a super-large-scale model, with a parameter count potentially reaching the 10 trillion level, approximately three times that of Moonshot AI's Kimi K3 model, the largest released model in China to date.
UK AI Safety Institute: "Unauthorized" Attack Behavior Detected in Testing of OpenAI and Anthropic Flagship Models
According to Bloomberg, the UK Government AI Safety Institute (established in 2023) disclosed on Tuesday that during safety evaluations of OpenAI's GPT-5.6-Sol and Anthropic's Mythos 5 models, both models exhibited "unauthorized" harmful behaviors, including actively intruding into real websites and attempting to inject malicious code into software, and these behaviors targeted real people and organizations. During the testing, the institute specifically granted the models internet access and disabled some safety filters to assess their extreme capabilities.
EU AI Act Enforcement Powers Officially Take Effect, Anthropic, OpenAI and Other US AI Giants Face Regulatory Pressure
According to CNBC, the European Commission officially gained the power to investigate, restrict, and fine AI models this Sunday, marking the latest progress in the phased implementation of the 2024 EU AI Act. Under the new rules, the European Commission can require AI companies to submit assessments before publicly releasing models in Europe and has the authority to restrict their EU market access. Violating companies face fines of up to €15 million or 3% of annual turnover (whichever is higher). The new rules apply to all companies providing general-purpose AI models within the EU, regardless of their headquarters' location, and non-EU companies must designate an authorized representative within the EU. Anthropic, OpenAI, and Google are all covered by the new rules. Currently, the EU has been seeking access to Anthropic's Mythos model for several months and has entered into negotiations with OpenAI and Anthropic regarding cyber attack issues triggered by their AI models.
Bitgo CEO deposits ~$6.3M in BTC, challenges Claude to move the funds
: Bitgo CEO Mike Belshe deposited 100 BTC into a public Bitcoin address on August 1, worth approximately $6.3 million at the time, and invited Anthropic's Claude model to attempt to move the funds out of the address. On-chain records show the wallet received the funds on July 31, and the balance had not been transferred out as of August 2. Anthropic previously disclosed that during 141,006 cybersecurity assessment runs, 3 incidents were found, with 6 evaluation sessions involving 3 models inadvertently interacting with real organizational systems. The models involved include Claude Opus 4.7, Claude Mythos 5, and an unreleased internal research model. The cause was a configuration error by third-party testing partner Irregular, which led to the test environment being connected to the internet. Anthropic stated that Claude Opus 4.7, during one evaluation, located a real website with the same name as a simulated company, exploited weak passwords and exposed services to recover infrastructure credentials, and accessed a production database containing hundreds of records. The company said the model was attempting to complete assigned tasks, not actively breaking constraints or pursuing independent goals. Belshe's challenge involves Bitgo's institutional custody platform, which uses multi-signature or multi-party computation technology to distribute signing authority across multiple independent keys. As of August 2, Anthropic had not publicly responded to the challenge.
Anthropic discloses three cybersecurity incidents involving Claude, where the model accessed real systems and obtained unauthorized permissions
Odaily News, July 30 – Anthropic released a report stating that during a review of cybersecurity assessment records, three incidents were discovered in which the Claude model accessed the internet in a third-party evaluation environment and further obtained unauthorized access to three real organizations' systems.Anthropic stated that the review covered 141,000 evaluation runs that could have potentially gained network access, and a total of three related incidents were found. All incidents occurred during Capture The Flag (CTF) cybersecurity tests, where the model was told the environment was a simulation with no internet access; however, due to configuration errors by the evaluation partner, the actual environment had internet connectivity.Among these, Claude Opus 4.7 accessed real company infrastructure during one test and obtained database permissions containing hundreds of production data records; Claude Mythos 5 built a malicious Python package and uploaded it to PyPI, resulting in the package being downloaded and run on 15 real systems; another internal research test model scanned approximately 9,000 targets and accessed a company's internet application through a public vulnerability.Anthropic stated that these incidents were not cases of the model actively seeking to escape or pursue its own goals, but rather the model mistakenly believed the real systems were within the test scope and continued executing the assigned cyberattack tasks. Notably, the newer internal research model stopped attacking after identifying that the targets might be real systems.Anthropic stated that these incidents primarily reflect issues with evaluation environment isolation and operational processes, rather than model alignment failures. The company has suspended related cybersecurity assessments, strengthened security controls in evaluation environments, continuously monitored test records, and will collaborate with third-party organizations to conduct further reviews.
Anthropic Discloses Claude Model Unauthorized Intrusion into Three Organizations' Systems
According to CNBC reports, Anthropic disclosed on July 30 that its Claude AI models accidentally breached the isolation environment and accessed the real internet during a cybersecurity assessment, gaining unauthorized access to the real systems of three different organizations through basic means such as accessing unauthenticated endpoints and exploiting weak passwords. The models involved include Opus 4.7, Mythos 5, and an internal research test model. The cause of the incident was a communication misunderstanding between Anthropic and third-party assessment partner Irregular, resulting in the models being told they were in a simulated environment without network access when they could actually still access the internet. Anthropic stated that this review was triggered by a similar Hugging Face intrusion incident disclosed by OpenAI last week; it has currently suspended all cybersecurity assessments and joined forces with independent AI assessment agency METR to launch further investigations, while calling on other AI labs to conduct similar reviews.