English translation is being prepared.
Matched fieldsTitle
←Search results
Search “Anthropic”
English translation is being prepared.
Matched fieldsTitle
English translation is being prepared.
Matched fieldsTitle
English translation is being prepared.
Matched fieldsTitle
English translation is being prepared.
Matched fieldsTitle
Anthropic studies behavioral tendencies and systemic failures of multi-agent systems
On August 13, 2026, Anthropic released a study on multi-agent systems that, through experiments including vulnerability discovery, collaborative game building, pricing games, and conflict transfer, identifies and demonstrates the behavioral tendencies of current frontier models and the unintended systemic failures they may cause. The study shows that a coordinated cluster of agents discovered 266 vulnerabilities within 27 million tokens, while independent parallel agents discovered 21 within 6.5 million tokens, with only 12 overlapping between the two. The study also found that agent behavior has low variance and agents readily make mistakes simultaneously, can rapidly form price collusion, and deploy malicious code against each other when goals conflict. The study is currently released by Anthropic as an official report, with no subsequent developments yet.
On August 13, 2026, Anthropic released a study on multi-agent systems that, through experiments including vulnerability discovery, collaborative game building, pricing games, and conflict transfer, identifies and demonstrates the behavioral tendencies of current frontier models and the unintended systemic failures they may cause. The study shows that a coordinated cluster of agents discovered 266 vulnerabilities within 27 million tokens, while independent parallel agents discovered 21 within 6.5
Matched fieldsTitle
Anthropic shares evidence review of worker retraining programs
On August 12, 2026, Anthropic shared an evidence review of worker retraining programs. The review was conducted by Anthropic together with independent researcher David Roodman, based on 56 U.S. randomized studies and European experimental evidence. The review found that job training programs have limited average effects: for each training slot provided, employment rates rise by 2 to 3 percentage points and annual earnings increase by about $1,000, while the cost is about $13,000; the government recovers more than half of the spending, leaving the program roughly break-even overall. "Sector programs" that partner with employers in high-demand industries and place workers directly in jobs yield several times greater returns, but attempts to replicate them often fail. The authors argue that if AI replaces jobs on a large scale, existing retraining programs may struggle to cope, and they recommend investing now in demonstration, evaluation, and scaling up the most promising programs. There are currently no subsequent developments.
On August 12, 2026, Anthropic shared an evidence review of worker retraining programs. The review was conducted by Anthropic together with independent researcher David Roodman, based on 56 U.S. randomized studies and European experimental evidence. The review found that job training programs have limited average effects: for each training slot provided, employment rates rise by 2 to 3 percentage points and annual earnings increase by about $1,000, while the cost is about $13,000; the government
Matched fieldsTitle
Anthropic studies behavioral tendencies and systemic failures in multi-agent systems
On August 13, 2026, Anthropic released a study on multi-agent systems that, through experiments including vulnerability discovery, collaborative game building, pricing games, and conflict transfer, identifies and demonstrates the behavioral tendencies of current frontier models and the unintended systemic failures they may cause. The study shows that a coordinated cluster of agents discovered 266 vulnerabilities within 27 million tokens, while independent parallel agents discovered 21 within 6.5 million tokens, with only 12 overlapping between the two. The study also found that agent behavior has low variance and agents readily make mistakes simultaneously, can rapidly form price collusion, and deploy malicious code against each other when goals conflict. The study is currently released by Anthropic as an official report, with no subsequent developments yet.
Matched fieldsTitle
View story:Anthropic studies behavioral tendencies and systemic failures of multi-agent systems
Anthropic shares evidence review of worker retraining programs
On August 12, 2026, Anthropic shared an evidence review of worker retraining programs. The review was conducted by Anthropic together with independent researcher David Roodman, based on 56 U.S. randomized studies and European experimental evidence. The review found that job training programs have limited average effects: for each training slot provided, employment rates rise by 2 to 3 percentage points and annual earnings increase by about $1,000, while the cost is about $13,000; the government recovers more than half of the spending, leaving the program roughly break-even overall. "Sector programs" that partner with employers in high-demand industries and place workers directly in jobs yield several times greater returns, but attempts to replicate them often fail. The authors argue that if AI replaces jobs on a large scale, existing retraining programs may struggle to cope, and they recommend investing now in demonstration, evaluation, and scaling up the most promising programs. There are currently no subsequent developments.
Matched fieldsTitle
View story:Anthropic shares evidence review of worker retraining programs
Anthropic uses Claude Mythos Preview to discover weaknesses in cryptographic algorithms
Anthropic researchers used Claude Mythos Preview to improve attacks on the post-quantum signature scheme HAWK, halving the effective key strength and reducing the key recovery cost for small-size HAWK-256 from 2^64 to 2^38. Another result targeted a 7-round reduced version of AES-128, where Claude proposed a method called Möbius Bridge, making the attack 200 to 800 times faster than the previous best meet-in-the-middle attack. Anthropic stated that neither result affects existing production systems: HAWK remains only a NIST third-round candidate, and the AES attack does not apply to the full 10-round encryption. Anthropic said each result cost about $100,000 in API costs, and that it built the CryptanalysisBench benchmark in collaboration with ETH Zurich, Tel Aviv University, and Technische Universität Berlin.
Anthropic researchers used Claude Mythos Preview to improve attacks on the post-quantum signature scheme HAWK, halving the effective key strength and reducing the key recovery cost for small-size HAWK-256 from 2^64 to 2^38. Another result targeted a 7-round reduced version of AES-128, where Claude proposed a method called Möbius Bridge, making the attack 200 to 800 times faster than the previous best meet-in-the-middle attack. Anthropic stated that neither result affects existing production syst
Matched fieldsTitle
Anthropic and Andon Labs develop drone control demonstration and evaluation
Anthropic and Andon Labs collaborate on Project Pilot, developing demonstrations and evaluations for assessing AI models' autonomous drone control. The task requires a drone to locate and track designated people indoors, and is split into five subtasks: reconstruction, localization, navigation, detection, and following, on which basis the benchmark Drone-Bench is built, with Andon Labs evaluating 15 models. The best performer is Claude Fable 5, which exceeds the human-AI team collaboration baseline on all subtasks except reconstruction, but because reconstruction errors accumulate in localization and navigation, it still cannot autonomously fly across rooms. Anthropic notes that the model capability frontier leads its stable performance by about six months, and Fable 5's current average level roughly corresponds to the one-off best performance of the previous-generation model in early 2026.
Anthropic and Andon Labs collaborate on Project Pilot, developing demonstrations and evaluations for assessing AI models' autonomous drone control. The task requires a drone to locate and track designated people indoors, and is split into five subtasks: reconstruction, localization, navigation, detection, and following, on which basis the benchmark Drone-Bench is built, with Andon Labs evaluating 15 models. The best performer is Claude Fable 5, which exceeds the human-AI team collaboration basel
Matched fieldsTitle
Anthropic uses Claude Mythos Preview to discover weaknesses in cryptographic algorithms
Anthropic researchers used Claude Mythos Preview to improve attacks on the post-quantum signature scheme HAWK, halving the effective key strength and reducing the key recovery cost for small-size HAWK-256 from 2^64 to 2^38. Another result targeted a 7-round reduced version of AES-128, where Claude proposed a method called Möbius Bridge, making the attack 200 to 800 times faster than the previous best meet-in-the-middle attack. Anthropic stated that neither result affects existing production systems: HAWK remains only a NIST third-round candidate, and the AES attack does not apply to the full 10-round encryption. Anthropic said each result cost about $100,000 in API costs, and that it built the CryptanalysisBench benchmark in collaboration with ETH Zurich, Tel Aviv University, and Technische Universität Berlin.
Matched fieldsTitle
View story:Anthropic uses Claude Mythos Preview to discover weaknesses in cryptographic algorithms
Anthropic and Andon Labs develop drone control demonstration and evaluation
Anthropic and Andon Labs collaborate on Project Pilot, developing demonstrations and evaluations for assessing AI models' autonomous drone control. The task requires a drone to locate and track designated people indoors, and is split into five subtasks: reconstruction, localization, navigation, detection, and following, on which basis the benchmark Drone-Bench is built, with Andon Labs evaluating 15 models. The best performer is Claude Fable 5, which exceeds the human-AI team collaboration baseline on all subtasks except reconstruction, but because reconstruction errors accumulate in localization and navigation, it still cannot autonomously fly across rooms. Anthropic notes that the model capability frontier leads its stable performance by about six months, and Fable 5's current average level roughly corresponds to the one-off best performance of the previous-generation model in early 2026.
Matched fieldsTitle
View story:Anthropic and Andon Labs develop drone control demonstration and evaluation
English translation is being prepared.
Matched fieldsTitle
English translation is being prepared.
Matched fieldsTitle
English translation is being prepared.
Matched fieldsTitle