AI, machine learning and LLM insights.
The update introduces two environment-variable caps – one for sub-agent spawns and one for web-search calls – both defaulting to 200 per session, with a 2-minute auto-background rule for long MCP calls, giving teams a concrete lever to curb runaway spend.
Noma Labs demonstrated that an attacker can post a crafted issue in a public repo, trigger a CI-linked AI agent with read rights to private repos, and have the bot dump those files back to the issue—all without stealing credentials or exploiting GitHub itself.
In real deployments every user utterance, tool schema, API response and retrieved document accumulates in the prompt. That hidden bloat makes processing time and pricing grow quadratically, turning a smooth demo into a costly latency nightmare.
The Linux Foundation’s new X402 Foundation, backed by Visa, Mastercard, Stripe and Google, defines the message format for AI-agent payments but omits any conformance suite, security profile, or certification process, leaving the authority claim untested.
Analyzing 312 real calls, the team replaced a 700 ms silence rule with a lightweight state machine that watches silence duration, grammatical completeness, and back-channel cues, driving truncations down from 18 % to 3 % and eliminating dead silence.
The benchmark, built on 163 labeled incidents, expands beyond diagnosis to test whether AI assistants can recognise uncertainty and deliberately abstain, highlighting a crucial safety gap as LLM confidence rises.
By repurposing models trained on unrelated image datasets, LIGO’s new system not only classifies known glitch families with near-perfect precision but also clusters novel anomalies, freeing scientists to focus on astrophysics instead of manual data cleaning.
In a single evening the agent delivered an MCP client, an Azure AI Agent, and an M365 Copilot Agent, but 12 out of 30 human messages were needed to steer it back, exposing workflow-violation and context-retrieval bugs that required prompt rewrites.
Google’s Gemini Notebook now indexes PDFs, Docs, and other files for instant AI queries, while Anthropic’s Claude can fetch credentials from 1Password to sign into services, and a joint spec aims to let agents discover and invoke APIs without custom code.
AWS Continuum’s AI agents continuously monitor version-control repos, suggest fixes and can even apply patches directly in developers’ IDEs or CI pipelines, slashing the window for supply-chain attacks.
The shift forces Google to redesign the entire chip stack, from layout to testing, and puts Intel’s yield capabilities under the microscope. Success could diversify the AI-chip supply chain; any hiccup may delay shipments and let TSMC reclaim dominance.
The modular pods, built in factories and shipped as complete units, can be connected in weeks, while Lancium’s software acts like a virtual battery, smoothing demand and tapping excess renewable or gas generation to keep the local ERCOT grid stable.
The operation serves polished videos to real users while showing bland pages to bots, using services like Keitaro to mask URLs. Experts suggest a verification pipeline that compares facial landmarks against trusted images to flag the fakes.
The new framework replaces a massive system prompt with a tiny SKILL.md file and loads policies or scripts only when the agent selects the right skill, slashing token usage, inference cost, and keeping reasoning crisp.
Production bot INKA-01 detected a hidden key item in the artwork, generated a detailed rejection, and routed the file back to design bot LUDO-01, preventing the error from reaching print without any human intervention.
DPO replaces the three-model RLHF pipeline with a single binary-cross-entropy loss on preference pairs, eliminating the reward model and PPO updates. The result is a stable, faster training loop that still hits or exceeds PPO benchmarks.
Anthropic has moved the Claude Fable 5 shutdown to 11:59 p.m. PT on July 19 2026 for all Pro, Max, Team and Enterprise tiers, while Claude Code users retain a 50 % weekly limit boost. Export-control reviews and security patches caused earlier slips.
Enterprise users can now upload a single selfie and voice clip to generate a lifelike digital twin, then harness Gemini Omni’s multimodal prompts to edit backgrounds, lighting and effects without restarting renders.
The feature, now in public alpha for verified New Zealand users, stitches proprietary Roblox models with open-source generative AI, producing environments, mechanics, assets and sound from a single text prompt that creators can then fine-tune before publishing.
The fintech team wrapped a thin reverse-proxy around Bedrock, vault-locking the real IAM credentials and issuing short-lived, per-team tokens that enforce model whitelists and hard token limits, delivering an audit trail and instant budget protection.
Governor Kathy Hochul fed large-language models every state rule, regulation and policy, prompting the AI to flag antiquated statutes—from a $25 dog-hunting fee to a midnight-work permit for pregnant employees—compressing a five-year task into a few months.
The EU’s Digital Markets Act compels Google to let any AI assistant—ChatGPT, Claude, Perplexity or others—access the same hardware hooks, sensors and app-level APIs it currently reserves for its Gemini model, with compliance due by July 2027.
A VentureBeat Pulse survey of 107 firms shows that while 83% run GPUs at half capacity or less, only 44% can accurately track compute spend, prompting 64% to consider switching or adding a cloud provider within the next year.
GPT-Red simulates dozens of attack vectors directly in developers’ CI/CD pipelines, surfacing vulnerabilities before a model ships. OpenAI says the tool trims the safety feedback loop from weeks to minutes, though it still advises human oversight for nuanced threats.
