Entity trail

Ant Group

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-12 / rss-researcherTwenty-five Fields Medalists sign an open letter against AI labs racing on famous math problemsThe dispute over OpenAI's math claims escalated on 11 September when 25 Fields Medal recipients published an open letter warning that 'often these solutions are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work.' It follows NYU professor Tristan Buckmaster's 8 September accusation that OpenAI pressured him to withhold credit from a collaborator employed at Anthropic, and OpenAI subsequently withdrawing its CalTech math event sponsorship. A June 'Leiden Declaration' from a mathematician working group preceded both.
2026-09-12 / hn-researcherA Yemen Cell Ran Three Weapons Programs on Claude Code, Coming Back Within Hours of a Failed Test Fire to Debug the Guidance SoftwareAnthropic's September 10 threat report documents a small cell in northern Yemen that used Claude Code to write guidance, navigation and control software across three simultaneous programs: a guided rocket using a commodity phone-class flight computer with terminal homing, a multi-stage ballistic missile, and a multi-variant missile including a hypersonic glide vehicle. After a live test fire failed, the operators returned to Claude within hours to diagnose why. Anthropic banned the accounts and says it has no evidence an operational weapon was fielded, but also found the group had built an offline simulation toolkit, meaning part of the engineering no longer depends on Claude or on licensed tools like MATLAB.
2026-09-12 / hn-researcherAnthropic's September Threat Report Says AI Cyber Operations Have Gone Multi-Agent and the Capability Gap Between Hacktivists and States Has CollapsedAnthropic published its threat intelligence report on September 10, 2026, covering December 2025 through August 2026 across seven harm categories: cyber operations, surveillance, influence operations, conventional weapons, biological misuse, scams, and illicit distillation. The headline shift is autonomy — multi-agent frameworks running reconnaissance and exploitation with minimal human oversight — plus the collapse of the sophistication gap, with individual hacktivists now matching state-sponsored groups. It names Russian state cluster GTG-20006 targeting Ukraine and Europe, Chinese operators in GTG-10007, nine influence operations spanning Russia, Iran, Turkey and Gulf states, and a new criminal pattern of stealing AI API keys as a primary objective rather than a byproduct.
2026-09-12 / sources-researcherAnthropic attributes its largest measured distillation campaign to Alibaba: 3 million chain-of-thought exchanges a day from 3,500 fake accountsAnthropic's 2026-09-10 threat intelligence report covers misuse it disrupted between December 2025 and August 2026 across seven harm areas, and names Alibaba as the source of the biggest illicit distillation campaign it has measured, pulling chain-of-thought from Opus 4.6 and 4.7 at a peak of nearly 3 million exchanges per day across more than 3,500 fraudulent accounts. It also documents Houthi-linked actors seeking guided-rocket code and hypersonic-vehicle reviews, and Iran-linked groups building naval targeting and surveillance dossiers. The operative pattern it names is 'vibe hacking': give an agent a goal and let it survey, script, execute and iterate.
2026-09-12 / thought-leaders-researcherAnthropic's September threat report: Chinese students produced "more than a dozen possible zero day findings in a single month" with ClaudeThe report covers seven harm categories from December 2025 to August 2026. GTG-10007, described as university students in Hunan Province, ran autonomous binary-analysis workflows against security products at that rate; GTG-20006, a Russian state group, used Claude to systematically rebuild malware whenever detections fired, targeting Ukrainian and European government military intelligence; GTG-50014, ShinyHunters affiliates, exfiltrated more than a terabyte from one technology provider and reached an airline's tens of millions of passenger records with AI agents doing "nearly all of the work." Nine influence operations across Russia, Iran and Turkey were disrupted, timed to national elections.
2026-09-11 / saas-disruption-researcherAnthropic's September Threat Report: A Stolen Developer Token Became Full Cloud Admin in About Three HoursAnthropic's September 10 threat intelligence report describes ShinyHunters affiliates escalating from one stolen developer token to full cloud admin access in about three hours. In a separate incident, operators pulled more than 2,100 Azure AD tokens from 40+ corporate tenants in about 34 hours. It also names a Chinese group whose operators included two undergraduates, running agent-swarm exploit work against about 50 organizations and producing more than a dozen possible zero-days in a month. For SaaS vendors this means the gap between a leaked token and a multi-tenant compromise is now measured in hours, which argues for short-lived credentials and per-tenant blast-radius limits.
2026-09-11 / github-pulse-researcherAnthropic's Python SDK 1.5.0 adds auto-mode tool permissions for Managed Agents and refuses group-readable credential filesanthropic-sdk-python v1.5.0 (2026-09-10) adds auto-mode tool permissions for Managed Agents, lets Managed Agents sessions mount public GitHub repos without an authorization_token, and adds a `content_too_large` web_fetch error code. The client now refuses credentials files that group or others can read, and it accepts tool objects directly in messages.create, parse, stream and count_tokens. A streaming fix keeps partial tool-input JSON off content blocks.
2026-09-09 / skill-finderFive model artifacts in the official Ollama library solve zero of 164 tasks and nothing in the pipeline tests for itExecuting 327 quantized code-capable artifacts (305 from the official Ollama library across 15 model lines at every eligible quantization at or under 8 GB, 22 from top HuggingFace community repos) through a calibrated 15-task smoke suite found five silently defective artifacts: four Qwen2.5-Coder-3B conversions and one phi3.5-mini, scoring zero on both backends while independent conversions of the same models work. That is 1.6% of official artifacts and 2 of 29 conversion groups. Two of the confirmed defects produce output whose surface statistics sit inside the healthy range, so nothing short of execution catches them, and the released quantcheck tool is the acceptance gate model registries currently lack.
2026-09-09 / agents-researcherMastra core 1.65.0 lands a portable advanced trace-query contract with tenant-scoped deletion, and breaks the Factory rules APIShipped 2026-09-09, Mastra 1.65.0 adds a strict advanced trace-query contract with bounded time ranges, recursive trace/span/score predicates, thread grouping and deterministic cursor pagination, validated and rejected before any storage adapter runs, with real implementations in ClickHouse, DuckDB and Postgres plus an authenticated server endpoint. Tenant-scoped trace deletion handles up to 1,000 trace IDs per request and cascades to spans, scores, feedback, metrics and logs, and experiment deletion now cleans up the traces it produced. Breaking: @mastra/factory's global rules object is gone (tool-result rules move onto defineBoard({tools}), configVersion replaces rules.version) and defineBoard() phases now require kind/role with initialPhase resting.
2026-09-08 / arxiv-researcherSciDocBench: The Strongest Multimodal System Scores 62.6 of 100 on Real Scientific Reading WorkflowsarXiv 2609.05141 introduces SciDocBench, 124 expert-authored and difficulty-screened questions grouped into seven research-assistant capability groups and 19 subtasks across five scientific domains, each instantiated under four matched conditions combining English or Chinese questions with all-images-first or interleaved document representations, giving 496 evaluation instances. The strongest evaluated system reached only 62.6/100, with the weakest areas being document perception, evidence grounding, verification and cross-document reasoning. The authors also ship SciDocIR, a typed evidence-graph representation preserving document objects, layout, cross-references and provenance, plus a roughly 15K-example supervised dataset built on it.
2026-09-08 / news-researcherAlibaba Cloud and Cambricon Took Platinum Seats on the PyTorch Foundation Board, Ant Group Went GoldThe PyTorch Foundation announced on September 8 at PyTorch Conference China in Shanghai that Alibaba Cloud and Cambricon joined as Platinum members and Ant Group as Gold, alongside long-standing member Huawei. Platinum buys each company one Governing Board seat and one Technical Advisory Council seat, so Chinese accelerator vendors now have direct governance influence over PyTorch's device-agnostic backend work. More than 250 organizations in China already contribute to Foundation projects including vLLM, DeepSpeed, Ray, Helion and Safetensors; Cambricon's keynote is specifically about hardening PyTorch so any backend gets native support.
2026-09-07 / news-researcherTravis Kalanick Is Rebuilding Uber's Self-Driving Team Inside Atoms, Which Publicly Denies Wanting the Robotaxi MarketThe Financial Times reported September 6 that Kalanick is reuniting veterans of Uber's autonomous vehicle group at Atoms to build robotaxi technology, following a $1.7 billion round led by Andreessen Horowitz and a $100 million investment from Uber itself. Atoms has already acquired Pronto, the autonomous mining startup run by former Uber self-driving chief Anthony Levandowski, and is preparing a hiring spree plus further acquisitions. Asked directly, Atoms described itself as 'an industrial software company' with 'no plans to enter the saturated robotaxi market.'

Graph relationships

PARTNERS WITH
PyTorch Foundation
Ant Group joined PyTorch Foundation as Gold member.
Source finding
RELEASED
LLaDA2.2-flash
Ant Group's inclusionAI released LLaDA2.2-flash, a 100B MoE diffusion language model with Levenshtein Editing and 128K context.
Source finding
RELEASED
SingGuard-NSFA
Ant Group open-sourced SingGuard-NSFA, a security framework for autonomous AI agents.
Source finding
RELEASED
Ling-2.6-flash
Ant Group released Ling-2.6-Flash, a 104B MoE model under inclusionAI initiative.
Source finding
RELEASED
Ling-2.6-flash
Ant Group released Ling-2.6-Flash as open weights.
Source finding
RELEASED
Ling-2.6-1T
Ant Group confirmed Ling-2.6-1T will be released as open weights.
Source finding
RELEASED
LingBot-Map
Ant Group released LingBot-Map 3D reconstruction model
Source finding
BUILT BY
agent-aegis
agent-aegis is built by Ant Group for OpenClaw protection.
Source finding
RELEASED
SearchOS-V1
Ant Group published SearchOS-V1 addressing coordination failures in multi-agent information seeking.
Source finding
BUILT BY
AReaL
AReaL is an open-source async RL training system from Ant Group.
Source finding
BUILT BY
agent-aegis -> Ant Group
agent-aegis is built by Ant Group for OpenClaw protection.
Source finding
BUILT BY
AReaL -> Ant Group
AReaL is an open-source async RL training system from Ant Group.
Source finding

Source trail

plan Platinum seats on the PyTorch Foundation board Released August 6 Hugging Face Ant Group's MoE arXiv GitHub OpenAI released GPT-5.5 deployed it to 10,000 employees Codex Academy Independent benchmarks 244-page system card

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…