Archive

623 stories in AI

Wednesday, 2 September 2026

A Closer Look at the AI Adoption Stages Behind Enterprise AI Success

Enterprises progress through four AI adoption stages: scattered tool use, initial pilots, live production, and shared business-wide systems. McKinsey's 2025 Global Survey found that 88 percent of companies have reached some level of AI adoption, though progression varies significantly across organizations.

Stolen Claude session cookies can reach corporate Gmail through grants no IT admin can revoke

Stolen Claude session cookies were replayed into paid accounts without triggering two-factor authentication, bypassing SSO protections on self-serve card-billed accounts. Anthropic identified six infostealer malware families, signed out affected accounts, removed saved payment methods, and refunded charges.

Google DeepMind Releases Gemini 3.8 Flash and Gemini 3.8 Flash Cyber: One Core Model, Two Access Envelopes

Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, sharing the same foundational model but differentiated by safety controls. Flash Cyber achieves 47.2% pass@1 on CWE-Bench and is restricted to vetted defenders through the Fairwind Program.

Pangram’s Max Spero on why AI detection is harder than ‘Real or Fake’

AI-generated content now appears in job applications, product reviews, and insurance claims, creating verification challenges across multiple sectors. Startups have emerged to address detection needs, though the article excerpt cuts off before explaining specific detection methods or capabilities.

US government sides with OpenAI on issue of training LLMs on copyrighted material

The U.S. government filed a brief supporting OpenAI's position that training large language models on copyrighted material constitutes fair use. The government stated it has a strong interest in developing a competitive AI industry that sets global standards for AI practice.

Motional and MIT AI explains self-driving car decisions

A self-driving car system developed by Motional and MIT researchers can explain its decisions in real-time, addressing the black-box problem in autonomous vehicle AI. The work was published in Nature and involved researchers from Motional and MIT's Computer Science and Artificial Intelligence Laboratory.

Forward-deployed engineering is how enterprise AI learns

Forward-deployed engineers embedded with customers can either become a sustainable product-learning function that generates reusable capabilities, or simply accumulate as delivery labor without creating lasting competitive advantage. The distinction determines whether subsequent customers begin with more developed product features and fewer unknowns, or whether vendors must restart with new services teams.

Meta Superintelligence Labs Releases Muse Voice Transcribe: One Real-Time Model for Streaming ASR, Diarization, and Endpointing

Meta released a single model that combines automatic speech recognition, speaker diarization, and endpointing detection instead of using three separate systems. The unified autoregressive model handles transcription, speaker separation, and user silence detection in one real-time streaming process.

Perplexity Releases Hybrid Compute on Mac: Cloud Agents Orchestrate Down to a Local Model, Gated On Device

Perplexity released hybrid compute for Mac that splits agent tasks between cloud models and local models, allowing sensitive data to remain on device. The system enables cloud-based agents to orchestrate decisions while delegating processing to a local model that handles privileged files and confidential information.

Tuesday, 1 September 2026

NVIDIA and CrowdStrike Strengthen Agentic Cybersecurity Frontier

CrowdStrike SafeMind, an agentic cybersecurity system, was announced as a collaboration between NVIDIA and CrowdStrike. The system uses automated defense capabilities to counter automated cyberattacks through agent-based technology.

Frontier models can recover up to 65% of facts they can't directly recall — just by thinking longer

Frontier language models can recover up to 65% of facts they fail to initially recall by using extended inference-time computation. Models like GPT-5 and Gemini-3 encode 95-98% of tested facts in their parameters, indicating recall rather than knowledge encoding is the primary bottleneck for factual accuracy.

Anthropic Announces Enterprise Frontier Safeguards, Customer-Held Data

Anthropic launched Enterprise Frontier Safeguards on September 1, 2026, which stores customer activity data in client-controlled cloud infrastructure instead of Anthropic servers while maintaining automated misuse detection. The system begins phased rollout in fall 2026 and applies to Mythos-class models like Claude Fable 5.1.

John Deere Puts JD AI Assistant Into Operations Center

John Deere embedded an AI assistant called JD into its Operations Center platform starting September 1, 2026, allowing farmers to ask plain-language questions about their field, machine, and operational data. The tool launched through an Early Access Program for select U.S. agricultural customers at the Farm Progress Show in Iowa.

AI is redefining the workforce — and most planning models aren’t ready

Sixty-two percent of C-suite executives report dissatisfaction with integration between people and business performance data across fragmented HR, finance, and procurement systems. Only 21% of organizations planning for AI's impact are addressing job design and organizational structure changes, while 50% focus solely on productivity and capacity effects.

Gradium AI Releases New Default TTS Model: 81.0% Hard-Case Pass Rate at 216 ms Time-to-First-Audio

A new text-to-speech model achieves an 81.0% human-rated pass rate on difficult sentences across five languages while delivering audio in 216 milliseconds. The evaluation used 500 hard-case sentences and the results are publicly available on Hugging Face under CC BY 4.0 licensing.

Sunday, 30 August 2026

AI agents need their own identity before they need a gateway

AI agents are shifting from answering questions to autonomously completing multi-step business workflows by dynamically deciding which tools to use and APIs to call. Once authenticated, traditional security controls lack visibility into whether agents continue operating safely, requiring enterprises to adopt runtime trust monitoring beyond initial identity verification.

Lowest-Latency Inference APIs for Voice and Realtime Agents: A Time to First Token TTFT-First Benchmark

A benchmark measured time to first token latency across voice agent stack layers including LLM, speech-to-text, text-to-speech, and speech-to-speech systems, with data verified August 30, 2026. Each latency figure was labeled as independently measured, vendor-published, or vendor-measured on its own product.

Google AI Introduces EnvHarness: A Programmable Layer That Turns Static Agent Environments Into Adaptive Training Worlds

A programmable layer called EnvHarness adapts static agent benchmarks during policy training while preserving original tasks and verifiers through standard interfaces. Across five benchmarks, mined skills improved by up to 9.0 points on held-out tasks with 9.8% fewer execution steps.

AI agents that pass authentication can still drift, expose data, or get memory-poisoned

# Summary A LiteLLM vulnerability exploited in June allowed attackers to run host commands through the gateway without credentials, representing one of seven CVEs disclosed in that AI gateway within a month. Enterprise teams prioritizing gateways as their first security control lack the underlying identity and attribution layers needed, making gateway deployment premature without established control planes that track agent identity, delegation, task assignment, and credential usage.

What Is Agentic AI? How Systems Plan, Use Tools, and Complete Tasks

Agentic AI systems pursue goals through repeated cycles of planning, tool use, observation, and adaptation rather than simply generating single answers. These systems operate in loops where they assess situations, execute actions using available tools, observe results, and adjust their approach accordingly.

Anthropic Opens a Research Preview of the Model Hardware Standard (MHS): A Shared Specification for AI Agents to Safely Operate Physical Devices

A shared driver specification called the Model Hardware Standard lets AI agents discover and safely operate physical devices by enforcing safety limits in the driver. Carnegie Mellon completed instrument integration in eight hours instead of weeks, and QuEra's laser relock accuracy improved from 58% to 99.3% across 700 trials.

Meet ‘Code-as-World’: An Agentic Loop That Rewrites Real Videos Into Executable MuJoCo Physics Programs

A system called Code-as-World extracts editable MuJoCo physics simulation code from real-world videos, then uses these reconstructed environments to train models on physical reasoning. The extracted scene code can be modified and executed within the MuJoCo simulator framework.

Friday, 28 August 2026

Anthropic Reports Claude Agents Mitigated Ten Alignment Failures

Claude agents autonomously developed training methods that mitigated ten alignment failures in target models while improving benchmarks without degrading general capabilities. Anthropic published this research on August 28, 2026, describing automated alignment post-training as potentially practical in the near term.

Blue Owl Funds Lead $2.4B AI Factory Equipment Financing for IREN

Blue Owl Capital is leading a $2.4 billion financing package for IREN Limited to purchase NVIDIA Blackwell Ultra GPUs for a data center in British Columbia. The financing consists of a $1.2 billion senior secured term loan and $1.2 billion in senior secured notes.

The three layers of agentic AI security: A defense-in-depth architecture for autonomous agents

Autonomous agents that can reason and execute actions across environments require security controls beyond traditional application-level protections. A defense-in-depth architecture spanning infrastructure, storage, compute, and networking layers is needed to address distinct risk categories including agent hallucinations, unauthorized credential use, and accidental data destruction.

Meta researchers taught an 8B AI model to match Claude Opus 4.5 — without the frontier price tag

Meta researchers developed EvoHarness-RL, a framework that trains an 8-billion-parameter AI model to perform at the level of Claude Opus 4.5. The system adds an abstraction layer to an agent's runtime environment, enabling the model to autonomously weigh costs and benefits rather than following rigid developer-written instructions.

← Prev1…345…21Next →

Get feedd. daily

Top stories in your inbox every morning. Pick what you want.

No spam. Unsubscribe anytime.