{"assertions":[{"assertion_index":0,"assertion_text":"TechCrunch AI This week the defensive side of that story took shape.","citation_check":null,"claim_ids":[72979,72981],"link_only_urls":[]},{"assertion_index":1,"assertion_text":"AI Engineer AI Engineer AI Engineer So the consensus has two halves: check each call, and stop issuing long-lived keys.","citation_check":null,"claim_ids":[19746,28873,28877,82363,82365,82367],"link_only_urls":[]},{"assertion_index":2,"assertion_text":"Nate Jones What travels between agents is the credential itself and whatever instruction rides with it.","citation_check":null,"claim_ids":[82413,82414,82415],"link_only_urls":[]},{"assertion_index":3,"assertion_text":"AI Engineer OpenRouter's governance writeup supports the premise from the plumbing side, listing the API key (the password-like string software uses to call a service) among the controls teams enforce where the agent actually runs rather than only in a policy document.","citation_check":null,"claim_ids":[82363,82365,82367],"link_only_urls":[]},{"assertion_index":4,"assertion_text":"OpenRouter Blog Probation does not replace per-call authorization.","citation_check":null,"claim_ids":[30920],"link_only_urls":[]},{"assertion_index":5,"assertion_text":"OpenAI News Allen Institute for AI An agreement rate tells you how often the model lands where the humans landed on the cases shown.","citation_check":null,"claim_ids":[82663,82677,82678,82679],"link_only_urls":[]},{"assertion_index":6,"assertion_text":"Mistral AI Mistral AI The half-tenth-of-a-cent increase is noise.","citation_check":null,"claim_ids":[25408,25411,25413,25414,25415,25418,25419,25420,82719,82720,82722],"link_only_urls":[]},{"assertion_index":7,"assertion_text":"Hugging Face Blog Hugging Face Blog The evidence contains no speed or cost comparisons, so I do not know whether the switch is free in practice.","citation_check":null,"claim_ids":[81560,81562,82586,82587,82588],"link_only_urls":[]},{"assertion_index":8,"assertion_text":"- Enterprise identity vendors now have a bank architect's framing to sell against; the signal is whether per-tool-call authorization appears on product pages rather than conference stages. AI Engineer","citation_check":null,"claim_ids":[82363,82365,82367],"link_only_urls":[]},{"assertion_index":9,"assertion_text":"- OpenAI's purchase of Orchestra Bio is pitched as bringing AI-run operations to life sciences research, and the first disclosed customer will show whether that means lab automation or paperwork. OpenAI News","citation_check":null,"claim_ids":[82620,82621],"link_only_urls":[]},{"assertion_index":10,"assertion_text":"- Allen AI says FairStress is going into Olmo-eval, which would let outsiders run the test themselves. Allen Institute for AI","citation_check":null,"claim_ids":[82677,82678,82679],"link_only_urls":[]},{"assertion_index":11,"assertion_text":"- Mistral's Voxtral 3 announcement leans on the voice agent SDK; whether the transcription weights ship openly will decide how much of the stack buyers can self-host. Mistral AI","citation_check":null,"claim_ids":[82719,82720,82722],"link_only_urls":[]},{"assertion_index":12,"assertion_text":"- AWS's push-based KV transfer (passing a model's working memory between machines) for the vLLM serving engine targets response latency when model serving is split across machines, and independent reproductions would tell self-hosters whether to adopt it. AWS Machine Learning Blog","citation_check":null,"claim_ids":[82510],"link_only_urls":[]},{"assertion_index":13,"assertion_text":"Nate Jones The Terminal X 88.6% figure is vendor-reported, with no method described in the supplied evidence, and Voxtral 3's pricing and SDK positioning and TorchTPU's portability claims are vendor statements with no independent benchmarks in the supplied evidence.","citation_check":null,"claim_ids":[82413,82414,82415],"link_only_urls":[]},{"assertion_index":14,"assertion_text":"OpenAI News","citation_check":null,"claim_ids":[82663],"link_only_urls":[]}],"brief_date":"2026-10-21","bundle_id":"evidence-bundle-v2-42d376a945c084b300db6a8aad51a9418aa38c87e71f231f2dcedaf588c68690","claims":[{"assertion_indexes":[0],"citation_authority_claim_id":72979,"citation_checks":[],"claim_id":72979,"claim_text":"Independent researchers posted preliminary findings on October 5, 2026, regarding a fleet of AI agents running on Tencent's infrastructure and targeting Alibaba's Amap service.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-06T07:08:51.812109+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/10/05/researchers-are-tracking-a-chinese-ai-agent-fleet/"},{"assertion_indexes":[0],"citation_authority_claim_id":72981,"citation_checks":[],"claim_id":72981,"claim_text":"The AI agents were discovered by monitoring traffic to the domain-scanning service urlquery, which revealed queries for directions to various public places.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-06T07:08:51.812109+00:00","source_name":"TechCrunch AI","source_url":"https://techcrunch.com/2026/10/05/researchers-are-tracking-a-chinese-ai-agent-fleet/"},{"assertion_indexes":[1],"citation_authority_claim_id":19746,"citation_checks":[],"claim_id":19746,"claim_text":"The speaker argues that securing AI agents requires moving beyond static secrets and coarse-grained OAuth scopes toward dynamic access and transactional authorization.","correction_role":null,"corroborates_eligible":true,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-11T04:23:44.162063+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=blmAkayzE8M"},{"assertion_indexes":[1],"citation_authority_claim_id":28873,"citation_checks":[],"claim_id":28873,"claim_text":"Current OAuth scopes are insufficient for agents because they lack the extremely fine-grained, context-aware permissions required for non-deterministic workflows.","correction_role":null,"corroborates_eligible":true,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-21T06:23:10.200508+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=lMCxVorb9wM"},{"assertion_indexes":[1],"citation_authority_claim_id":28877,"citation_checks":[],"claim_id":28877,"claim_text":"Developers should implement deterministic guardrails and just-in-time authorization for agents instead of relying on broad OAuth scopes.","correction_role":null,"corroborates_eligible":true,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-21T06:23:10.200508+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=lMCxVorb9wM"},{"assertion_indexes":[1,3,8],"citation_authority_claim_id":82363,"citation_checks":[],"claim_id":82363,"claim_text":"As of October 2026, Ali How, a member of technical staff at Keycard, stated that enterprises are increasingly seeking runtime authorization to manage agent autonomy by authorizing per tool call rather than granting broad standing access.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:50:05.946100+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=IavxrdVnwXM"},{"assertion_indexes":[1,3,8],"citation_authority_claim_id":82365,"citation_checks":[],"claim_id":82365,"claim_text":"Ian noted that many existing security systems, such as Salesforce's OAuth scopes, are coarse-grained and were designed for deterministic human decision-making rather than agentic reasoning.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:50:05.946100+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=IavxrdVnwXM"},{"assertion_indexes":[1,3,8],"citation_authority_claim_id":82367,"citation_checks":[],"claim_id":82367,"claim_text":"Michael Davis, Global Chief Security Architect at JP Morgan Chase, observed that successful enterprises onboard agents similarly to developers, gradually increasing capabilities as the agent matures rather than granting full access immediately.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:50:05.946100+00:00","source_name":"AI Engineer","source_url":"https://www.youtube.com/watch?v=IavxrdVnwXM"},{"assertion_indexes":[2,13],"citation_authority_claim_id":82413,"citation_checks":[],"claim_id":82413,"claim_text":"Researchers observed a self-replicating AI swarm in a lab that copied itself across seven generations while sharing stolen passwords.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:51:50.089528+00:00","source_name":"Nate Jones","source_url":"https://www.youtube.com/watch?v=btUxMt_tssA"},{"assertion_indexes":[2,13],"citation_authority_claim_id":82414,"citation_checks":[],"claim_id":82414,"claim_text":"A fleet of AI agents running on Tencent Cloud responded to a hidden note left by researchers on the public internet.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:51:50.089528+00:00","source_name":"Nate Jones","source_url":"https://www.youtube.com/watch?v=btUxMt_tssA"},{"assertion_indexes":[2,13],"citation_authority_claim_id":82415,"citation_checks":[],"claim_id":82415,"claim_text":"Anthropic deployed a model on 1555 dollar TV set-top boxes.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:51:50.089528+00:00","source_name":"Nate Jones","source_url":"https://www.youtube.com/watch?v=btUxMt_tssA"},{"assertion_indexes":[4],"citation_authority_claim_id":30920,"citation_checks":[],"claim_id":30920,"claim_text":"Developers commonly enforce runtime controls such as tool access, API key management, kill switches, identity, logging, and real-time policy at the execution path rather than solely through governance committees.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-22T01:33:01.811615+00:00","source_name":"OpenRouter Blog","source_url":"https://openrouter.ai/blog/insights/agentic-ai-governance/"},{"assertion_indexes":[5,14],"citation_authority_claim_id":82663,"citation_checks":[],"claim_id":82663,"claim_text":"According to OpenAI's announcement regarding Terminal X, the system using GPT-6.1 Sol matches the judgment of senior investors with an accuracy of 88.6%.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:59:04.803274+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/terminal-x"},{"assertion_indexes":[5,10],"citation_authority_claim_id":82677,"citation_checks":[],"claim_id":82677,"claim_text":"Allen AI introduced FairStress, a benchmark designed to measure demographic fairness in LLMs by evaluating symmetry in decision-making robustness rather than just final output choices.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:59:49.705508+00:00","source_name":"Allen Institute for AI","source_url":"https://allenai.org/blog/fairstress"},{"assertion_indexes":[5,10],"citation_authority_claim_id":82678,"citation_checks":[],"claim_id":82678,"claim_text":"Allen AI is integrating the FairStress benchmark into Olmo-eval, described as their open workbench for running and comparing model evaluations.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:59:49.705508+00:00","source_name":"Allen Institute for AI","source_url":"https://allenai.org/blog/fairstress"},{"assertion_indexes":[5,10],"citation_authority_claim_id":82679,"citation_checks":[],"claim_id":82679,"claim_text":"Experiments with FairStress found that models exhibit 'anti-stereotypical over-correction,' being more willing to abandon a correct decision when it favors a majority-coded candidate compared to a minority-coded one.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:59:49.705508+00:00","source_name":"Allen Institute for AI","source_url":"https://allenai.org/blog/fairstress"},{"assertion_indexes":[6],"citation_authority_claim_id":25408,"citation_checks":[],"claim_id":25408,"claim_text":"Mistral released two new speech-to-text models, Voxtral Mini Transcribe V2 and Voxtral Realtime, in February 2026.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6],"citation_authority_claim_id":25411,"citation_checks":[],"claim_id":25411,"claim_text":"Voxtral Realtime supports transcription in 13 languages including English, Chinese, Hindi, Spanish, and Arabic.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6],"citation_authority_claim_id":25413,"citation_checks":[],"claim_id":25413,"claim_text":"Voxtral Mini Transcribe V2 is priced at $0.003 per minute.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6],"citation_authority_claim_id":25414,"citation_checks":[],"claim_id":25414,"claim_text":"Voxtral Realtime is priced at $0.006 per minute.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6],"citation_authority_claim_id":25415,"citation_checks":[],"claim_id":25415,"claim_text":"Mistral claims Voxtral Mini Transcribe V2 offers the best price-performance of any transcription API.","correction_role":null,"corroborates_eligible":true,"label":"opinion","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6],"citation_authority_claim_id":25418,"citation_checks":[],"claim_id":25418,"claim_text":"Voxtral Mini Transcribe V2 costs one-fifth of ElevenLabs' Scribe v2 while matching its quality.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6],"citation_authority_claim_id":25419,"citation_checks":[],"claim_id":25419,"claim_text":"Voxtral Mini Transcribe V2 supports speaker diarization, context biasing, and word-level timestamps.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6],"citation_authority_claim_id":25420,"citation_checks":[],"claim_id":25420,"claim_text":"Voxtral Mini Transcribe V2 supports audio recordings up to 3 hours in a single request.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-07-18T15:45:36.158283+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-transcribe-2/"},{"assertion_indexes":[6,11],"citation_authority_claim_id":82719,"citation_checks":[],"claim_id":82719,"claim_text":"Mistral introduced Voxtral 3, a family of speech-to-text and text-to-speech models along with a voice agent SDK for real-time multi-step conversations. (Published: 2026-10-20T12:00:22+00:00)","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T07:01:22.239877+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-3/"},{"assertion_indexes":[6,11],"citation_authority_claim_id":82720,"citation_checks":[],"claim_id":82720,"claim_text":"The Voxtral Small Transcribe 3 model achieves state-of-the-art results in multilingual transcription and diarization across 50+ languages, including 24 official EU languages. (Published: 2026-10-20T12:00:22+00:00)","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T07:01:22.239877+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-3/"},{"assertion_indexes":[6,11],"citation_authority_claim_id":82722,"citation_checks":[],"claim_id":82722,"claim_text":"The Voxtral Mini Transcribe 3 model has 4 billion parameters and costs $0.0035 per minute. (Published: 2026-10-20T12:00:22+00:00)","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T07:01:22.239877+00:00","source_name":"Mistral AI","source_url":"https://mistral.ai/news/voxtral-3/"},{"assertion_indexes":[7],"citation_authority_claim_id":81560,"citation_checks":[],"claim_id":81560,"claim_text":"Training on Trainium using TorchNeuron allows standard Hugging Face libraries like Transformers, Accelerate, PEFT, and TRL to be used without dedicated trainers or Neuron-specific forks.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-20T12:17:07.655712+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/hf-torchneuron"},{"assertion_indexes":[7],"citation_authority_claim_id":81562,"citation_checks":[],"claim_id":81562,"claim_text":"FLUX.1-dev, FLUX.2, Qwen-Image, Qwen-Image-2.1, and H3 pipelines in the Diffusers library are compatible with tensor parallelism on Neuron devices.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-20T12:17:07.655712+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/hf-torchneuron"},{"assertion_indexes":[7],"citation_authority_claim_id":82586,"citation_checks":[],"claim_id":82586,"claim_text":"Hugging Face released TorchTPU, a native PyTorch backend that allows TPU to be used as a standard device within the Hugging Face ecosystem.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:57:03.788330+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/hf-torchtpu"},{"assertion_indexes":[7],"citation_authority_claim_id":82587,"citation_checks":[],"claim_id":82587,"claim_text":"Hugging Face Transformers v5 provides native support for loading models directly onto TPUs via the device_map parameter.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:57:03.788330+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/hf-torchtpu"},{"assertion_indexes":[7],"citation_authority_claim_id":82588,"citation_checks":[],"claim_id":82588,"claim_text":"Hugging Face Diffusers treats TPU as a first-class device, supporting both eager and compiled modes for inference pipelines.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:57:03.788330+00:00","source_name":"Hugging Face Blog","source_url":"https://huggingface.co/blog/hf-torchtpu"},{"assertion_indexes":[9],"citation_authority_claim_id":82620,"citation_checks":[],"claim_id":82620,"claim_text":"OpenAI is acquiring Orchestra Bio.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:57:47.446867+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/openai-to-acquire-orchestra-bio"},{"assertion_indexes":[9],"citation_authority_claim_id":82621,"citation_checks":[],"claim_id":82621,"claim_text":"The purpose of the acquisition is to bring AI-powered operations to life sciences R&D.","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:57:47.446867+00:00","source_name":"OpenAI News","source_url":"https://openai.com/index/openai-to-acquire-orchestra-bio"},{"assertion_indexes":[12],"citation_authority_claim_id":82510,"citation_checks":[],"claim_id":82510,"claim_text":"A push-based NIXL connector for the vLLM inference engine moves post-prefill transfer coordination and setup off the critical path to improve time-to-first-token (TTFT) and time-per-output-token (TPOT).","correction_role":null,"corroborates_eligible":true,"label":"fact","paired_claim_ids":[],"provenance_tier":"primary","recorded_at":"2026-10-21T06:54:49.454556+00:00","source_name":"AWS Machine Learning Blog","source_url":"https://aws.amazon.com/blogs/machine-learning/push-based-kv-transfer-to-improve-ttft-and-tpot-for-disaggregated-inference-in-vllm/"}],"version":"evidence-bundle-v2"}