{"assertions":[{"assertion_index":0,"assertion_text":"Epoch AI My view is more urgent: today's evidence shows that old weaknesses become dangerous when agents erase the labor cost of exploiting them.","citation_check":null,"claim_ids":[35136],"link_only_urls":[]},{"assertion_index":1,"assertion_text":"Princeton CITP The model did not invent the flaw.","citation_check":null,"claim_ids":[41646,41647,41648],"link_only_urls":[]},{"assertion_index":2,"assertion_text":"Import AI That is not an outbreak.","citation_check":null,"claim_ids":[41548,41550,41551],"link_only_urls":[]},{"assertion_index":3,"assertion_text":"Embrace The Red An IT leader should treat routing credentials as production secrets and keep agent permissions narrow even when the underlying model is trusted.","citation_check":null,"claim_ids":[41603,41604],"link_only_urls":[]},{"assertion_index":4,"assertion_text":"Microsoft Research OpenRouter separately launched Ori Eval to compare models on a builder's own prompts and data.","citation_check":null,"claim_ids":[41611,41612,41614],"link_only_urls":[]},{"assertion_index":5,"assertion_text":"OpenRouter Orchard supplies exactly the kind of headline benchmark number an engineer should re-run on their own tasks, and Ori Eval is a tool for doing that comparison.","citation_check":null,"claim_ids":[41493],"link_only_urls":[]},{"assertion_index":6,"assertion_text":"ChinaAI The capability pressure has been building since mid-July, and a newer comparison put DeepSeek R1 only a few points behind leading closed models, displacing an earlier concern that it could not keep up.","citation_check":null,"claim_ids":[41516,41517,41518],"link_only_urls":[]},{"assertion_index":7,"assertion_text":"AI Engineer One Useful Thing A narrower capability gap does not make the hardware burden disappear.","citation_check":null,"claim_ids":[471,15736,15741,15742],"link_only_urls":[]},{"assertion_index":8,"assertion_text":"OpenAI A business leader should not treat that as a transferable return.","citation_check":null,"claim_ids":[41489,41490,41491],"link_only_urls":[]},{"assertion_index":9,"assertion_text":"- Georgia election officials should publish any scanner or records-format changes that prevent casting-order reconstruction before the next public export. Princeton CITP","citation_check":null,"claim_ids":[41646,41647,41648],"link_only_urls":[]},{"assertion_index":10,"assertion_text":"- LiteLLM maintainers should add public guidance on admin-route controls, provider-key isolation, and tool-call logging. Embrace The Red","citation_check":null,"claim_ids":[41603,41604],"link_only_urls":[]},{"assertion_index":11,"assertion_text":"- Independent teams need to reproduce Orchard's coding and browser results, including the cost of ranking multiple answers. Microsoft Research","citation_check":null,"claim_ids":[41611,41612,41614],"link_only_urls":[]},{"assertion_index":12,"assertion_text":"- Moonshot AI's next useful signal is an official smaller deployment profile with measured power and operating cost. ChinaAI","citation_check":null,"claim_ids":[41516,41517,41518],"link_only_urls":[]},{"assertion_index":13,"assertion_text":"- Circles should disclose cohort definitions and measurement periods behind its revenue and customer-loss gains. OpenAI","citation_check":null,"claim_ids":[41489,41490,41491],"link_only_urls":[]},{"assertion_index":14,"assertion_text":"Princeton CITP Import AI Microsoft Research OpenAI ChinaAI","citation_check":null,"claim_ids":[41489,41490,41491,41516,41517,41518,41548,41550,41551,41611,41612,41614,41646,41647,41648],"link_only_urls":[]}],"brief_date":"2026-08-04","bundle_id":"evidence-bundle-v1-c532c566fd46552e8fab86ed8ea7837f5596ba84666a63c7d575d775daecd07b","claims":[{"assertion_indexes":[0],"citation_authority_claim_id":35136,"citation_checks":[],"claim_id":35136,"claim_text":"Multiple independent benchmarks, including ExploitGym, ExploitBench, UK AI Security Institute's Cyber Ranges, Irregular's CyScenarioBench and FrontierCyber, have shown frontier AI models can discover security vulnerabilities in real-world code and develop exploits to hack realistic systems.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-23T01:04:07.949674+00:00","source_name":"Epoch AI Gradient Updates","source_url":"https://epochai.substack.com/p/openai-accidentally-hacked-hugging"},{"assertion_indexes":[1,9,14],"citation_authority_claim_id":41646,"citation_checks":[],"claim_id":41646,"claim_text":"A 2022 research report identified a critical privacy vulnerability in certain ballot scanning machines where the algorithm used to assign random numbers to electronic ballot records was deterministic and reversible, allowing the sequence of ballot casting to be reconstructed.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:56:22.513259+00:00","source_name":"Princeton CITP - Freedom to Tinker","source_url":"https://blog.citp.princeton.edu/2026/08/03/an-algorithmic-failure-beneath-the-secret-ballot/"},{"assertion_indexes":[1,9,14],"citation_authority_claim_id":41647,"citation_checks":[],"claim_id":41647,"claim_text":"Author Max Springer used AI coding agents and public records to reconstruct the casting order of 1.52 million in-person ballots (98.9% of the total) from Georgia's May 2026 primary election.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:56:22.513259+00:00","source_name":"Princeton CITP - Freedom to Tinker","source_url":"https://blog.citp.princeton.edu/2026/08/03/an-algorithmic-failure-beneath-the-secret-ballot/"},{"assertion_indexes":[1,9,14],"citation_authority_claim_id":41648,"citation_checks":[],"claim_id":41648,"claim_text":"In 114 of the 139 Georgia counties examined, the in-person casting order of ballots could be recovered from public files alone, enabling the identification of voters' ballots in specific contexts.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:56:22.513259+00:00","source_name":"Princeton CITP - Freedom to Tinker","source_url":"https://blog.citp.princeton.edu/2026/08/03/an-algorithmic-failure-beneath-the-secret-ballot/"},{"assertion_indexes":[2,14],"citation_authority_claim_id":41548,"citation_checks":[],"claim_id":41548,"claim_text":"Researchers from the University of Toronto, the Vector Institute, the University of Cambridge, and ServiceNow built a prototype computer virus that uses AI models to compromise computers and leverages their GPU resources for inference to infect more hosts.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:54:23.245583+00:00","source_name":"Import AI","source_url":"https://importai.substack.com/p/import-ai-467-self-sustaining-ai"},{"assertion_indexes":[2,14],"citation_authority_claim_id":41550,"citation_checks":[],"claim_id":41550,"claim_text":"The proof-of-concept worm operates using an open-weight LLM published in 2025 that fits on a single A100 GPU with 80GB of VRAM, without relying on monitored vendor APIs.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:54:23.245583+00:00","source_name":"Import AI","source_url":"https://importai.substack.com/p/import-ai-467-self-sustaining-ai"},{"assertion_indexes":[2,14],"citation_authority_claim_id":41551,"citation_checks":[],"claim_id":41551,"claim_text":"The prototype worm achieved approximately 80% success in vulnerability detection, 53% in exploitation, and 88% in self-replication, resulting in an overall attack success rate of about 37%.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:54:23.245583+00:00","source_name":"Import AI","source_url":"https://importai.substack.com/p/import-ai-467-self-sustaining-ai"},{"assertion_indexes":[3,10],"citation_authority_claim_id":41603,"citation_checks":[],"claim_id":41603,"claim_text":"A malicious actor can hijack LiteLLM traffic, steal backend LLM provider keys, and inject tool calls by exploiting legitimate proxy-admin credentials and routing configuration endpoints.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:55:24.237407+00:00","source_name":"Embrace The Red","source_url":"https://embracethered.com/blog/posts/2026/hijacking-litellm-for-fun-and-profit/"},{"assertion_indexes":[3,10],"citation_authority_claim_id":41604,"citation_checks":[],"claim_id":41604,"claim_text":"In March 2026, the PyPI package for LiteLLM was compromised and distributed a credential-stealing payload.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:55:24.237407+00:00","source_name":"Embrace The Red","source_url":"https://embracethered.com/blog/posts/2026/hijacking-litellm-for-fun-and-profit/"},{"assertion_indexes":[4,11,14],"citation_authority_claim_id":41611,"citation_checks":[],"claim_id":41611,"claim_text":"Microsoft Research introduced Orchard, an open-source framework centered on Orchard Env, a Kubernetes-based environment service designed to support scalable agentic AI research across multiple task domains.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:55:35.711884+00:00","source_name":"Microsoft Research Blog","source_url":"https://www.microsoft.com/en-us/research/blog/orchard-an-open-framework-for-scalable-agentic-ai/"},{"assertion_indexes":[4,11,14],"citation_authority_claim_id":41612,"citation_checks":[],"claim_id":41612,"claim_text":"Orchard-SWE achieved a 69.7% score on the SWE-bench Verified benchmark using approximately 3 billion active parameters, reaching 73.0% with value-model reranking.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:55:35.711884+00:00","source_name":"Microsoft Research Blog","source_url":"https://www.microsoft.com/en-us/research/blog/orchard-an-open-framework-for-scalable-agentic-ai/"},{"assertion_indexes":[4,11,14],"citation_authority_claim_id":41614,"citation_checks":[],"claim_id":41614,"claim_text":"Orchard-GUI, a 4-billion-parameter vision-language model, achieved an average success rate of 68.4% across the WebVoyager, Online-Mind2Web, and DeepShop benchmarks.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:55:35.711884+00:00","source_name":"Microsoft Research Blog","source_url":"https://www.microsoft.com/en-us/research/blog/orchard-an-open-framework-for-scalable-agentic-ai/"},{"assertion_indexes":[5],"citation_authority_claim_id":41493,"citation_checks":[],"claim_id":41493,"claim_text":"OpenRouter launched Ori Eval, a tool designed to help users identify the best AI model for their specific application by running evaluations on the user's own prompts and data.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:47:47.970954+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/announcements/ori-eval/"},{"assertion_indexes":[6,12,14],"citation_authority_claim_id":41516,"citation_checks":[],"claim_id":41516,"claim_text":"Deploying Moonshot AI's K3 model requires at least 16 H200 GPUs, a hardware cost that makes personal deployment unaffordable for most users.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:48:12.507928+00:00","source_name":"ChinaAI","source_url":"https://chinai.substack.com/p/chinai-369-my-boss-wants-me-to-run"},{"assertion_indexes":[6,12,14],"citation_authority_claim_id":41517,"citation_checks":[],"claim_id":41517,"claim_text":"Moonshot AI recommends deploying K3 on a super-node of at least 64 accelerator cards, with total costs of approximately 17 million RMB.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:48:12.507928+00:00","source_name":"ChinaAI","source_url":"https://chinai.substack.com/p/chinai-369-my-boss-wants-me-to-run"},{"assertion_indexes":[6,12,14],"citation_authority_claim_id":41518,"citation_checks":[],"claim_id":41518,"claim_text":"Running 64 K3 accelerator cards at full load consumes 45 kilowatts of electricity, exceeding the capacity of ordinary household electrical meters and wiring.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:48:12.507928+00:00","source_name":"ChinaAI","source_url":"https://chinai.substack.com/p/chinai-369-my-boss-wants-me-to-run"},{"assertion_indexes":[7],"citation_authority_claim_id":471,"citation_checks":[],"claim_id":471,"claim_text":"DeepSeek r1 is missing features compared to other major AI companies and may not keep up in the long term.","correction_role":null,"label":"forecast","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-06-08T12:53:19.292361+00:00","source_name":"One Useful Thing","source_url":"https://www.oneusefulthing.org/p/using-ai-right-now-a-quick-guide"},{"assertion_indexes":[7],"citation_authority_claim_id":15736,"citation_checks":[],"claim_id":15736,"claim_text":"OpenAI's o3 model leads in intelligence, followed closely by o4 mini with high reasoning mode and Deepseek R1.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T02:52:09.555543+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=sRpqPgKeXNk"},{"assertion_indexes":[7],"citation_authority_claim_id":15741,"citation_checks":[],"claim_id":15741,"claim_text":"The gap between open-weight model intelligence and proprietary model intelligence has narrowed significantly, with Deepseek R1 being only a few points behind leading models.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T02:52:09.555543+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=sRpqPgKeXNk"},{"assertion_indexes":[7],"citation_authority_claim_id":15742,"citation_checks":[],"claim_id":15742,"claim_text":"Deepseek and Alibaba are the leading providers of open-weight reasoning and non-reasoning models.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T02:52:09.555543+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=sRpqPgKeXNk"},{"assertion_indexes":[8,13,14],"citation_authority_claim_id":41489,"citation_checks":[],"claim_id":41489,"claim_text":"Circles utilizes the OpenAI API and Codex to deliver AI-native telecommunications experiences.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:47:40.618532+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/circles"},{"assertion_indexes":[8,13,14],"citation_authority_claim_id":41490,"citation_checks":[],"claim_id":41490,"claim_text":"The implementation of OpenAI technology by Circles resulted in a 22% increase in average revenue per user (ARPU).","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:47:40.618532+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/circles"},{"assertion_indexes":[8,13,14],"citation_authority_claim_id":41491,"citation_checks":[],"claim_id":41491,"claim_text":"The implementation of OpenAI technology by Circles resulted in a 9% reduction in customer churn.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-04T06:47:40.618532+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/circles"}],"version":"evidence-bundle-v1"}