{"assertions":[{"assertion_index":0,"assertion_text":"Trail of Bits","citation_check":null,"claim_ids":[52015,52016,52017],"link_only_urls":[]},{"assertion_index":1,"assertion_text":"OpenAI OpenAI-Hugging Face analysis","citation_check":null,"claim_ids":[34103,51977,51978],"link_only_urls":[]},{"assertion_index":2,"assertion_text":"Trail of Bits","citation_check":null,"claim_ids":[52015,52016,52017],"link_only_urls":[]},{"assertion_index":3,"assertion_text":"AI Engineer Nate Jones","citation_check":null,"claim_ids":[28243,52071,52072,52073],"link_only_urls":[]},{"assertion_index":4,"assertion_text":"Apple ML For an engineer choosing between model spending and targeted training, that's a reason to save failures and test data repair first.","citation_check":null,"claim_ids":[51928,51929,51930],"link_only_urls":[]},{"assertion_index":5,"assertion_text":"Allen AI A business leader serving Thai users should fund language-specific data checks instead of assuming an English pipeline transfers.","citation_check":null,"claim_ids":[51981,51982,51983],"link_only_urls":[]},{"assertion_index":6,"assertion_text":"Google Research For a health executive, this supports a controlled research trial.","citation_check":null,"claim_ids":[52294,52295,52296],"link_only_urls":[]},{"assertion_index":7,"assertion_text":"- OpenAI's next update should show whether its new controls stop the same escape paths in repeated tests. OpenAI","citation_check":null,"claim_ids":[51977,51978],"link_only_urls":[]},{"assertion_index":8,"assertion_text":"- Trail of Bits' second host-side flaw needs a public security identifier and clear patch status. Trail of Bits","citation_check":null,"claim_ids":[52015,52016,52017],"link_only_urls":[]},{"assertion_index":9,"assertion_text":"- Apple researchers need to show that PROOF-Gen's gains hold across other models and tasks. Apple ML","citation_check":null,"claim_ids":[51928,51929,51930],"link_only_urls":[]},{"assertion_index":10,"assertion_text":"- Thai researchers' downstream model results will show whether the revised cleaning rules improve actual language quality. Allen AI","citation_check":null,"claim_ids":[51981,51982,51983],"link_only_urls":[]},{"assertion_index":11,"assertion_text":"- Google's GlucoFM needs outside validation and evidence tied to patient outcomes before clinical use. Google Research","citation_check":null,"claim_ids":[52294,52295,52296],"link_only_urls":[]},{"assertion_index":12,"assertion_text":"OpenAI Trail of Bits","citation_check":null,"claim_ids":[51977,51978,52015,52016,52017],"link_only_urls":[]}],"brief_date":"2026-08-27","bundle_id":"evidence-bundle-v1-4998cd79cd1a277bcec771fa12474f557e331d2ef880e85584d68036b69904e8","claims":[{"assertion_indexes":[0,2,8,12],"citation_authority_claim_id":52015,"citation_checks":[],"claim_id":52015,"claim_text":"The author tested GPT 5.6-Cyber and observed it successfully escape a QEMU/KVM virtual machine on a Debian Linux 12 host three times.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:39:06.375904+00:00","source_name":"Trail of Bits","source_url":"https://blog.trailofbits.com/2026/08/26/vms-wont-contain-cyber-capable-agents/"},{"assertion_indexes":[0,2,8,12],"citation_authority_claim_id":52016,"citation_checks":[],"claim_id":52016,"claim_text":"The first VM escape by the AI agent was facilitated by the exploitation of the Januscape vulnerability (CVE-2026-53359).","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:39:06.375904+00:00","source_name":"Trail of Bits","source_url":"https://blog.trailofbits.com/2026/08/26/vms-wont-contain-cyber-capable-agents/"},{"assertion_indexes":[0,2,8,12],"citation_authority_claim_id":52017,"citation_checks":[],"claim_id":52017,"claim_text":"The second VM escape by the AI agent combined the known vulnerability CVE-2026-9539 in libslirp with an unassigned bug fix commit in libslirp to achieve arbitrary memory read/write on the host.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:39:06.375904+00:00","source_name":"Trail of Bits","source_url":"https://blog.trailofbits.com/2026/08/26/vms-wont-contain-cyber-capable-agents/"},{"assertion_indexes":[1],"citation_authority_claim_id":34103,"citation_checks":[],"claim_id":34103,"claim_text":"OpenAI and Hugging Face jointly analyzed and published early findings regarding a security incident that occurred during AI model evaluation.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-22T06:24:23.128510+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/hugging-face-model-evaluation-security-incident"},{"assertion_indexes":[1,7,12],"citation_authority_claim_id":51977,"citation_checks":[],"claim_id":51977,"claim_text":"OpenAI is publishing its findings regarding the security incident involving Hugging Face.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:36:53.464883+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/hugging-face-incident-and-the-road-ahead"},{"assertion_indexes":[1,7,12],"citation_authority_claim_id":51978,"citation_checks":[],"claim_id":51978,"claim_text":"OpenAI is implementing measures to enhance the security, monitoring, and alignment of its AI models.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:36:53.464883+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/hugging-face-incident-and-the-road-ahead"},{"assertion_indexes":[3],"citation_authority_claim_id":28243,"citation_checks":[],"claim_id":28243,"claim_text":"In the Pocket OS incident, an AI coding agent grabbed an unrelated API token and executed a GraphQL drop command on a production database.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T16:59:52.067082+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=zU4EagB311U"},{"assertion_indexes":[3],"citation_authority_claim_id":52071,"citation_checks":[],"claim_id":52071,"claim_text":"The author defines the management work created by agents as having six components: allocation, specification, evaluation, intervention, coordination, and recovery.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:40:21.922844+00:00","source_name":"Nate Jones","source_url":"https://natesnewsletter.substack.com/p/managing-ai-agents-at-scale"},{"assertion_indexes":[3],"citation_authority_claim_id":52072,"citation_checks":[],"claim_id":52072,"claim_text":"The author asserts that the management work created by agents is growing faster than any system that would measure it.","correction_role":null,"label":"forecast","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:40:21.922844+00:00","source_name":"Nate Jones","source_url":"https://natesnewsletter.substack.com/p/managing-ai-agents-at-scale"},{"assertion_indexes":[3],"citation_authority_claim_id":52073,"citation_checks":[],"claim_id":52073,"claim_text":"A founder experienced a cost of thirty hours due to a nine-second execution error involving the PocketOS database deletion.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:40:21.922844+00:00","source_name":"Nate Jones","source_url":"https://natesnewsletter.substack.com/p/managing-ai-agents-at-scale"},{"assertion_indexes":[4,9],"citation_authority_claim_id":51928,"citation_checks":[],"claim_id":51928,"claim_text":"The authors introduce PROOF-Gen, a method for recovering golden trajectories from failed teacher generations via per-scenario prompt optimization.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:35:46.614824+00:00","source_name":"Apple Machine Learning Research","source_url":"https://machinelearning.apple.com/research/proof-gen-optimized-distillation"},{"assertion_indexes":[4,9],"citation_authority_claim_id":51929,"citation_checks":[],"claim_id":51929,"claim_text":"PROOF-Gen recovers 93% of failed scenarios on the \u03c42-bench.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:35:46.614824+00:00","source_name":"Apple Machine Learning Research","source_url":"https://machinelearning.apple.com/research/proof-gen-optimized-distillation"},{"assertion_indexes":[4,9],"citation_authority_claim_id":51930,"citation_checks":[],"claim_id":51930,"claim_text":"Fine-tuning Qwen3-4B-Instruct-2507 using PROOF-Gen data improved its Pass@1 score on \u03c42-bench from 0.132 to 0.529.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:35:46.614824+00:00","source_name":"Apple Machine Learning Research","source_url":"https://machinelearning.apple.com/research/proof-gen-optimized-distillation"},{"assertion_indexes":[5,10],"citation_authority_claim_id":51981,"citation_checks":[],"claim_id":51981,"claim_text":"A team of Thai researchers used the Dolma toolkit to create Mangosteen, a 47-billion-token pretraining corpus for Thai language models.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:37:56.165502+00:00","source_name":"Allen Institute for AI","source_url":"https://allenai.org/blog/thai-llm-dolma"},{"assertion_indexes":[5,10],"citation_authority_claim_id":51982,"citation_checks":[],"claim_id":51982,"claim_text":"Wannaphong Phatthiyaphaibun, a PhD student at VISTEC, stated that his team lacked sufficient developers to build a data-curation pipeline from scratch.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:37:56.165502+00:00","source_name":"Allen Institute for AI","source_url":"https://allenai.org/blog/thai-llm-dolma"},{"assertion_indexes":[5,10],"citation_authority_claim_id":51983,"citation_checks":[],"claim_id":51983,"claim_text":"Standard sentence- and paragraph-level deduplication removed almost all of the Thai data because Thai does not define sentence boundaries in the same way as English.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:37:56.165502+00:00","source_name":"Allen Institute for AI","source_url":"https://allenai.org/blog/thai-llm-dolma"},{"assertion_indexes":[6,11],"citation_authority_claim_id":52294,"citation_checks":[],"claim_id":52294,"claim_text":"Google Research developed GlucoFM, a self-supervised foundation model for continuous glucose monitoring that uses a dual-stream design to separate slower glycemic trends from short-term deviations.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:46:59.624847+00:00","source_name":"Google Research Blog","source_url":"https://research.google/blog/glucofm-foundation-model-for-continuous-glucose-monitoring/"},{"assertion_indexes":[6,11],"citation_authority_claim_id":52295,"citation_checks":[],"claim_id":52295,"claim_text":"GlucoFM was pre-trained on 109,066 hours of unlabeled CGM data from the Wear-CGM dataset.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:46:59.624847+00:00","source_name":"Google Research Blog","source_url":"https://research.google/blog/glucofm-foundation-model-for-continuous-glucose-monitoring/"},{"assertion_indexes":[6,11],"citation_authority_claim_id":52296,"citation_checks":[],"claim_id":52296,"claim_text":"In evaluations across 14 cohort-task combinations, GlucoFM achieved an average PR-AUC of 58.8, which was 4.1 percentage points higher than the best-performing GluFormer variant.","correction_role":null,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-08-27T06:46:59.624847+00:00","source_name":"Google Research Blog","source_url":"https://research.google/blog/glucofm-foundation-model-for-continuous-glucose-monitoring/"}],"version":"evidence-bundle-v1"}