Dakarda AI Newsletter · 7 August 2026
Friday Edition
AI Agent in the Enterprise: 20× Cheaper
The week when AI agents stopped being a developer experiment — Meta enters with the cheapest coding agent on the market, Microsoft gives every agent its own identity in the company, and OWASP warns against excessive agency.
The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).
Intro · Alex
I've been watching for some time now how AI agents are ceasing to be a developer toy and are becoming a real tool in enterprises. This week only confirmed it. Meta released Muse Code — a coding agent at 1/20th the price of competitors, but in exchange for allowing them to train on your code. Microsoft introduced Entra Agent ID — every bot in the company gets its own account with permissions and audit. And OWASP published the first-ever list of threats for agents based on real incidents, not surveys. In this edition, I've traced the most important moves of this week: from Meta's pricing strategy through agent identities at Microsoft, to the geopolitics of open models and the warning from Hugging Face that China is taking the lead. You'll also find a practical tip of the day — a technique that could protect your agent from the fastest-growing threat of 2026.
What's worth knowing
Meta Launches Muse Code — Terminal Coding Agent at 1/20th the Price of Competitors
Meta Superintelligence Labs (Alexander Wang's team acquired for ~$14.3B) released Muse Code — an AI agent running on macOS and Linux, powered by a dedicated Muse Spark 1.2 model. It launches multiple persistent agents in the background (planning, debugging, repo searching), with a standard price of $1.25/M input. A separate 'Contributor Tier' offers ~20× discount in exchange for allowing Meta to use your code for training their models.
OWASP GenAI Top 10 2026 — First Edition Based on Real Incidents, Not Surveys
The OWASP GenAI Security Project published two new lists: Top 10 for LLM Applications 2026 and Top 10 for Agentic Applications 2026 (new). For the first time, the ranking was based on data from real incidents. 'Excessive Agency' — too much agent autonomy — moved up to 3rd place, and 'Unbounded Consumption' (uncontrolled resource consumption) jumped 4 positions. Over 100 experts participated in the work.
Microsoft Entra Agent ID — First Identities for AI Agents in Enterprises
Microsoft introduced Entra Agent ID (public preview) — an extension of Microsoft Entra ID with identity constructs designed specifically for AI agents. Each agent gets an individual, policy-managed account with its own permissions, roles, and audit. Until now, agents operated on shared accounts — now each has its own identity, which is a breakthrough for compliance and risk management.
Hugging Face CEO: 'China Dominates Open Models — Could Lead by End of 2026'
Clément Delangue, CEO of Hugging Face, told CNBC that China 'clearly dominates in open models' and could take the lead on the AI frontier by the end of 2026 or 2027. He attributed this to the Chinese model of collaboration, while American companies 'build in silos.' Meanwhile, DeepSeek raised prices, and US startups are trying to catch up. For the entire developer ecosystem, this is a signal that model selection, API pricing, and regulations could change soon.
From the tech world
Thousands of Servers Can Be Backdoored via Vulnerable BMC Controllers
Researchers discovered five classes of critical vulnerabilities in BMC chips (including HPE iLO, Supermicro, OpenBMC, Dell). Flaws in IPMI handshake, unsigned firmware, and KCS access allow for persistent backdooring of servers. The article contains specific defensive recommendations: disable IPMI, KCS, and isolate BMCs on separate VLANs.
OWASP Top 10 for LLM 2026 — AI Agent Got Hands, Keys, and Access to Corporate Systems
A Polish overview of the updated OWASP Top 10 for LLM 2026. Excessive Agency moved from 6th to 3rd place, and Improper Output Handling dropped to 10th — but its scope was expanded to include dangerous AI-generated code. A practical guide for anyone deploying agents in their company.
Tip of the day
Principle of Least Privilege for AI Agents
In the latest OWASP Top 10, 'Excessive Agency' moved up to 3rd place — agents too often get more permissions than they need. The principle is simple: an agent should only have access to the systems and data necessary to perform a specific task, and only for its duration. This sounds obvious, but in practice, most deployments give agents full access 'just in case.' Practical step: Before deploying an agent, create a 'scope card' — a document with three columns: (1) tasks it needs to perform, (2) systems it needs access to, (3) allowed operations on each system (read/write/delete). Mark each excessive permission as a risk and justify it. This takes 15 minutes per agent and eliminates the biggest threat of 2026.
Tool of the issue
Muse Code — Meta's Coding Agent at 1/20th the Price of Competitors
A terminal AI agent from Meta running on macOS and Linux, powered by a dedicated Muse Spark 1.2 model. It launches multiple persistent agents in the background — planning, repository searching, debugging. Standard price: $1.25/M input tokens, with a Contributor Tier option (~20× cheaper) in exchange for allowing training on your code.
Reading list
Meta Launches Muse Code: A New AI Coding Agent Powered by Spark 1.2
A full analysis of Meta's strategy: why they're entering at such a low price, how the Contributor Tier works, and what it means for the coding agent market. Required reading if you're considering switching tools.
Thousands of Servers Can Be Backdoored by Exploiting Buggy Motherboard Controllers
A detailed list of five classes of BMC vulnerabilities with specific defensive recommendations. If you manage servers — this could be the most important security article you'll read this month.
When an AI agent gets its own corporate account and OWASP warns against excessive agency — it's a sign that the era of agents in the enterprise has truly begun.
Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.