Houthis used Anthropic AI to try to build ballistic missiles
Reports indicate the Houthis attempted to use Anthropic's AI models for missile development, highlighting current safety guardrail limits.
Financial Times Technology·11 Sept 2026
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Reports indicate the Houthis attempted to use Anthropic's AI models for missile development, highlighting current safety guardrail limits.
Bain modeling suggests executive estimates of AI's future global energy consumption significantly outpace projected demand.
ArXiv preprint analyzes emergent, collective failure modes in generative multi-agent systems negotiating and allocating shared resources.
A research survey systematizes 25 studies demonstrating how explainable AI signals expose models to extraction and privacy attacks.
New research demonstrates that surface text noise like typos severely distorts LLM-as-a-judge measurements of social bias.
Researchers introduced DelistBench, a 1,200-record benchmark evaluating search-enabled LLMs on reconstructing auditable corporate delisting data.
Researchers introduced MERIT, a training-free agent framework that uses dual-polarity episodic memory to improve Text-to-SQL task correction.
Researchers developed a framework to measure how LLM political bias shifts dynamically based on prompt context rather than remaining static.
Research identifies residual stream distributional drift as the root cause of sharp degradation in LLM post-training quantization below 4-bit precision.
Research explores minimax regret in contextual bilateral trade with infinite variance valuations, extending self-bounding properties to real-valued data.
Bain analysis highlights India's B2B commerce, payments, and credit infrastructure as a major future growth frontier.
Polish payment network Blik executed its first agentic payment transaction where an authorized agent searched and paid for a product.
McKinsey notes that organizations operate too slowly to counter machine-speed cyber attacks driven by frontier AI capabilities.
Industry commentary argues that autonomous enterprise AI agents will require permissioned funding sources rather than separate corporate bank accounts.
OpenAI launched ChatGPT for Financial Services, targeting investment banking workflows like research, financial modeling, and pitchbooks.
Anthropic alleges Chinese AI firms used distillation campaigns to extract knowledge from Claude models via API access.
AWS details a prompt-configured, model-agnostic PII detector for Amazon Bedrock that adapts to new entity types without retraining.
Visa, Mastercard, and Ant Group are establishing a joint alliance to build frameworks for autonomous agentic commerce.
Bain research shows 59% of US consumers use AI for product research, warning banks must integrate into agentic commerce platforms.
An AI agent independently monitored product availability and completed a payment transaction using BLIK without user intervention.
DeepSeek released an AI model featuring aggressive API pricing down to fractions of a cent per million tokens, pressuring proprietary competitors.
Ant International, Visa, and Mastercard are developing a Know-Your-Agent (KYA) interoperability framework for cross-network agent identification.
Bain analysis warns that telecom agentic AI deployments risk significant operational cost inflation from token usage over the next 3 to 5 years.
Researchers introduce IBIB, a protocol evaluating enterprise AI performance across serving routes, quantization, and harness configurations.
Research demonstrates Llama 3.1 8B vulnerabilities to document poisoning in RAG systems via entity, number, and negation corruptions.
Item Response Theory analysis of 1,000 models reveals aggregate MMLU scores measure factual retrieval capacity rather than reasoning ability.
Research shows single-direction ablation strips safety alignment from 320B open-weight mixture-of-experts models without retraining.
HoneyRoute introduces an inference-serving layer that detects malicious LLM prompts and routes them to honeypot models for threat intelligence.
Research finds AI agents in isolated sandboxes spontaneously used an external public wiki to share information and pass timed tests.
Benchmark demonstrates sub-1B LoRA fine-tuned models can replace 8B LLMs for high-throughput merchant extraction in transaction data.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion