AMD · Jul 22
AMD launches its Helios AI system and MI450 chips, and takes a $5B stake in Anthropic
At its Advancing AI 2026 event, AMD introduced the Helios rack-scale system built on Instinct MI455X GPUs and 2-nanometer EPYC Venice CPUs, plus the MI450 accelerator line. AMD and Anthropic agreed to deploy up to 2 gigawatts of MI450 GPUs starting in the first half of 2027, with AMD taking a strategic equity stake of up to $5 billion in Anthropic. AMD said OpenAI and Meta have now committed a combined 12 gigawatts of its accelerators.
▸ The MSP Angle
Does a stronger AMD change the price of the AI I run for clients?
A real number two to Nvidia is good news for anyone who buys compute, which is every AI service you run for clients. More supply and price competition at the chip layer is what eventually pulls per-token costs down and loosens the GPU scarcity that inflates AI pricing today. You will not switch silicon yourself, but you feel it in the rates your model providers charge. Build your delivery so it can ride falling compute costs into better margins rather than locking clients into today's prices.
Read at AMD ↗
The Hacker News · Jul 23
A Claude desktop-agent flaw could reach the whole hard drive on 500,000 Macs
Researchers at Accomplish AI disclosed SharedRoot, a sandbox escape in Anthropic's Claude Cowork app for macOS that mounted the entire host filesystem into the agent's virtual machine with read and write access. Chained with a kernel flaw, an AI agent could read or change any file on the Mac with no permission prompt, affecting roughly 500,000 users who ran local sessions. Anthropic closed the report as informative without a dedicated fix; current builds default to running in the cloud, but people who enabled local execution stay exposed.
▸ The MSP Angle
Are desktop AI agents a new endpoint risk I have to manage?
Yes. A desktop AI agent with local file access is a new class of endpoint software, and its default permissions can reach far wider than anyone reviewing the install would guess. This one could touch every file on the machine without a single prompt. Add AI agent apps to your endpoint inventory, check what they can reach on client devices, and prefer configurations that run in a controlled environment over ones that hand an agent the whole disk. This is managed-endpoint work now, not a curiosity.
Read at The Hacker News ↗
The Hacker News · Jul 24
AI pentest agents found eight serious flaws in a popular forum platform in six hours
Aikido Security said its AI pentest agents uncovered eight high-severity vulnerabilities in the NodeBB forum platform after six hours of automated source-code review. The flaws, in versions before 4.14.0 and fixed in 4.14.2, could expose the admin dashboard without a login, leak private messages and user lists, and run malicious code through spoofed federation servers; five of the eight sit in federation code that ships enabled by default. No CVE identifiers were assigned.
▸ The MSP Angle
How does AI-powered hacking change how fast I have to patch?
Faster than your current schedule assumes. When an AI can read a codebase and surface eight serious bugs in an afternoon, the gap between a release and a working exploit keeps shrinking, and it shrinks for attackers and defenders alike. Two moves: tighten patch windows on internet-facing client software, and turn AI-speed discovery to your own advantage by running the same kind of automated review across the apps you manage. The patch cadence you set last year is already too slow.
Read at The Hacker News ↗
Pax8 · Jul 13
Pax8: 61% of small businesses now actively use AI, but only 23% have a written AI policy
Pax8's Q2 2026 SMB AI Pulse Report, based on 402 U.S. small-business leaders, found 61% actively using AI and 29% experimenting, with the interested-but-not-started group collapsing to 1.5% from 9% a quarter earlier. AI users pull ahead of non-users: 65% feel very or extremely confident about growth versus 56%, and 53% raised technology spending versus 24%. Only 23% have a documented AI policy, and the top barriers are lack of internal expertise and unclear return.
▸ The MSP Angle
Most of my clients already use AI, so what do they still need from me?
The number that matters is not the 61% using AI, it is the 23% with a written policy and the crowd stuck on no expertise and unclear return. Adoption is finished; the gap now is doing it safely and provably. That gap is a managed service: an AI use policy, guardrails on what data goes where, and reporting that ties the tools to results. Clients who already spend more on technology once they adopt AI are telling you they will pay to close it.
Read at Pax8 ↗
AWS · Jul 20
AWS launches CloudWatch Coding Agent Insights to measure the ROI of AI coding tools
Amazon CloudWatch introduced Coding Agent Insights, which gives engineering leaders a view into how AI coding tools actually drive value. Built on OpenTelemetry metrics, it reads Claude Code telemetry with no added instrumentation and also supports Codex, letting teams see where agents speed delivery and right-size token budgets by department. It is available in all AWS commercial regions except the UAE, Bahrain, and Tel Aviv.
▸ The MSP Angle
How do I prove and bill the value of the AI tools I put in front of clients?
This is a template worth borrowing. The hardest part of selling AI is showing the return, and the answer here is boring in the best way: telemetry, per-team usage, and token budgets you can report on. Whether or not you touch this tool, wire measurement into every AI service you deliver so you can show a client where it saved hours, and set spend limits before a surprise bill lands. Provable return is what turns an AI pilot into a renewing line item.
Read at AWS ↗
ChannelE2E · Jul 22
Barracuda expands its MSP partner program with new marketing and certification tools
Barracuda expanded its Partner Success Program with co-branded email, social, and digital campaign tools, an upgraded partner portal, an enlarged Mastery Program that adds technical and sales badges, and a coming Partner Locator. The vendor framed the update around helping partners build pipeline and prove expertise as demand for AI governance grows.
▸ The MSP Angle
Are vendor partner programs actually useful for growing an AI practice?
The useful parts of a partner program are not the logo, they are the demand-generation and certification pieces you would otherwise build yourself. Co-branded campaigns fill a pipeline; badges give buyers a reason to trust you on AI security. Judge any program by whether it hands you working sales motion and credible proof of skill, and skip the ones that are only a discount sheet. As demand for AI governance rises, provable expertise is exactly what clients are shopping for.
Read at ChannelE2E ↗
TechCrunch · Jul 7
Norm AI raises $120M at a $1.2B valuation to build agents that enforce regulations
Norm AI closed a $120 million Series C led by Khosla Ventures at a $1.2 billion valuation, with Blackstone, Bain Capital Ventures, Coatue, and others joining and total funding passing $260 million in under three years. The company builds AI agents that read regulations, monitor compliance in real time, and police how other AI systems behave. Blackstone is both an investor and a major customer.
▸ The MSP Angle
Is there real money in AI compliance, or is it just overhead?
A billion-dollar valuation for agents that watch other agents tells you where this is heading: oversight of AI is becoming its own category, not a cost center. Your clients will soon need someone to answer what their AI did, why, and whether it followed the rules. That is a service you can offer now with far less capital than a startup needs, using policy, logging, and review on top of the tools clients already run. Position AI governance as a line you sell, not a favor you throw in.
Read at TechCrunch ↗
Amazon · Jul 9
OpenAI's GPT-5.6 model family reaches general availability on Amazon Bedrock
The full GPT-5.6 family reached general availability on Amazon Bedrock: Sol for advanced reasoning, Terra for production work at about half the cost of the prior generation, and Luna for fast, high-volume tasks. Bedrock adds hardware-enforced isolation, in-region processing, a 90% discount on cached input, pricing matched to OpenAI's own rates, and usage that counts toward existing AWS commitments, all under one API alongside models from Anthropic, Meta, Mistral, Cohere, and Amazon.
▸ The MSP Angle
Should I run client AI on one model, or keep my options open?
Keep your options open, and make it architectural. Governed access to many models under a single interface is what lets you move a client to a cheaper or better model the week it lands, with no rebuild and no new contract. A production tier at half the cost of last generation is margin only if your delivery can adopt it cleanly. Build for a world where the model underneath changes often and the client never has to care which one is running.
Read at Amazon ↗
OpenAI · Jul 22
OpenAI introduces Presence, a managed platform for deploying voice and chat agents
OpenAI launched Presence, a managed platform for building production voice and chat agents with built-in guardrails, policy enforcement, simulation-based testing, and post-launch tuning. OpenAI said Presence already runs its own English-language phone support line and resolves 75% of inbound calls with no human. It is available through limited general availability, with rollouts led by OpenAI's own engineers and a set of chosen systems integrators.
▸ The MSP Angle
What does it mean when the model makers start selling the deployment layer too?
It means the build-and-govern-the-agent layer is now contested, which is worth watching rather than fearing. A model maker shipping a managed agent platform, delivered through its own engineers and a few chosen integrators, signals how much value sits in deployment and not just the model. For MSP partners the takeaway is positioning: the moat is the client relationship, the industry context, and the accountability for what an agent does day to day, none of which ship in a platform. Note where the market is moving and make sure your role is the one a vendor's rollout team cannot take.
Read at OpenAI ↗
