Author: Amitabh Sarkar

Published: August 7, 2026 Human reviewers approved 1 in 3 malicious AI agent commands as routine across 40,000 simulated oversight sessions, according to research published August 5, 2026 by security researcher Alex Wauters at Scale X. The study analysed 409,000 individual approve/deny decisions and found a mean threat-detection accuracy of 66.3%, meaning the average human-in-the-loop reviewer missed 33.7% of threats presented to them. What the Scale X Study Found The study used a browser-based simulation at llmgame.scalex.dev where participants act as the human overseer for an AI coding agent, approving or denying terminal commands under time pressure. Approximately 34% of…

Published: August 7, 2026 Anthropic launched inference hooks in beta on August 5, 2026 — a feature for Claude Enterprise that routes every employee prompt through the organisation’s own security server for an allow-or-deny decision before Claude processes the message. The feature closes the compliance gap that left enterprise AI chat outside the data-loss-prevention (DLP) controls applied to email, web traffic, and file transfers. How Claude Enterprise Inference Hooks Work When an employee submits a prompt in Claude Enterprise chat, Claude Code, or Claude Cowork, the inference hook sends a signed HTTPS POST of the conversation transcript to the organisation’s…

Microsoft’s Service Agent inside Microsoft 365 Copilot now closes Dynamics 365 customer service cases, drafts resolution emails, and updates CRM records directly from Outlook and Teams, according to a July 15, 2026 feature breakdown published by the Microsoft Dynamics 365 team. Saurabh Gupta and Rushil Vora of the Microsoft Dynamics 365 team detailed the capabilities in a post titled “From Answers to Action: What’s New with Copilot in Dynamics 365 Customer Service.” Service Agent reached general availability on June 30, 2026; its counterpart Sales Agent followed on July 7. Both agents connect to live Dynamics 365 CRM data through a…

Atlassian’s Rovo AI assistant has an unpatched indirect prompt injection vulnerability that lets attackers silently exfiltrate Jira tickets, Confluence documents, and connected third-party data, security firm PromptArmor disclosed publicly on August 5, 2026 — more than two months after reporting the flaw to Atlassian on May 23. PromptArmor reported the vulnerability to Atlassian on May 23, 2026. Atlassian assigned a case number on May 25 and then made no further substantive communication despite PromptArmor follow-ups on June 4 and July 29. PromptArmor published full technical details on August 5 after exhausting its responsible disclosure process. Rovo is the AI assistant…

Published August 6, 2026 Meta launched Muse Code on August 5, 2026 — its first AI coding agent — alongside Muse Spark 1.2, a coding-focused model update, entering a market where Anthropic’s Claude Code and OpenAI Codex are the current leaders. Muse Code is a terminal agent for macOS and Linux (beta) that plans code changes, writes and validates them across large repositories, and installs with a single command. Standard pricing is $1.25 per million input tokens and $4.25 per million output tokens; a “contributor tier” cuts output cost to $0.20 per million — approximately 21 times cheaper — in…

Published August 6, 2026 Demis Hassabis stepped down as CEO of Google DeepMind on August 5–6, 2026, moving to Chair of Google DeepMind and Chief Scientist of Alphabet, while Jeff Dean — a 27-year Google veteran and co-founder of Google Brain — departs to co-found an AI research startup called Discovery Loop. Koray Kavukcuoglu, formerly DeepMind’s CTO and Alphabet’s Chief AI Architect, takes over as SVP of Google DeepMind. Alphabet shares fell approximately 4% following the announcement. What Changed at Google DeepMind on August 5, 2026? Google DeepMind announced 4 simultaneous leadership changes: Hassabis exits the CEO role, becoming Chair…

Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs launched Ode with Anthropic on July 15, 2026 — a $1.5 billion joint venture that embeds engineers inside enterprises to wire AI models into core business processes. The venture, first announced without a name in May 2026, was built on the acquisition of Fractional AI, a boutique AI engineering startup that previously spent 11 months as an OpenAI implementation partner. Ode employs approximately 100 engineers, more than half of them former founders. What Ode with Anthropic Does Ode with Anthropic is an AI implementation company: it deploys forward-deployed engineers (FDEs) who work…

Published: August 5, 2026 | Updated: August 2026 Microsoft Copilot Studio’s agentic workflow designer is now generally available, letting organizations build multi-step AI automations on a visual canvas — no custom code required. The August 2026 release adds multi-agent orchestration, evaluation automation APIs, and run-only agent sharing in public preview, backed by a real-world deployment at Graebel that processes over 50,000 service orders per year. Microsoft Copilot Studio is a low-code platform that lets business users build, deploy, and orchestrate AI agents and automated workflows inside the Microsoft 365 ecosystem without writing code. Microsoft announced the initial GA rollout in…

Cloudflare launched Cloudflare Wallets and cloudflare.pay on August 4, 2026, giving AI agents deployed on its infrastructure a persistent identity and a virtual wallet with human-set spending limits. Each Cloudflare account receives a unique web address that functions as a stable identity anchor for any AI agent running on it, and operators can issue per-agent “Virtual Wallets” with configurable spending caps, merchant allowlists, and maximum per-transaction amounts. The system uses stablecoins rather than traditional payment rails. How Cloudflare Wallets Work Cloudflare Wallets operates on a two-tier structure. The Account Wallet is the central balance — it can receive, hold, and…

Mistral AI released Shieldstral on August 4, 2026 — a 3B-parameter open-weights multimodal AI safety classifier that runs on a single 16GB NVIDIA GPU and accepts plain-language policy questions at inference time, requiring no retraining to add new moderation categories. The model is licensed under Apache 2.0 and is free to download and self-host. Shieldstral achieves an average F1 score of 84.9% across safety benchmarks and outperforms models up to 7× its size on text safety tasks, according to Mistral AI. The release was announced on August 4, 2026 via the Mistral AI blog. What Shieldstral Does Shieldstral classifies text…