9to5Mac · Aug 6
OpenAI publishes an open standard for packaging agent skills so they move between tools
OpenAI published the Agent Plugins v1.0.0 specification on August 6, an openly licensed format that bundles agent skills and Model Context Protocol servers into portable packages any compatible client can discover and load. It is developed in public with a multi-vendor steering committee that includes Amazon and Vercel, and reframes the GPT-5 anniversary week around interoperability rather than a new model.
▸ The MSP Angle
Will the AI agents I build for one client work anywhere, or am I locked in?
An open, multi-vendor packaging standard is the difference between an agent you build once and reuse across clients and one you rebuild every time a vendor changes. If the automations and skills your team creates become portable packages, the work you do for the first client becomes inventory you deploy to the next, which is how a repeatable AI service line gets its margin. Favor platforms that speak open standards like MCP and this one over closed ecosystems, because portability is what turns one-off builds into a practice.
Read at 9to5Mac ↗
Help Net Security · Aug 3
Horizon3.ai raises $250 million at a $2 billion valuation for autonomous penetration testing
Horizon3.ai closed a $250 million Series E co-led by NightDragon and NEA, tripling its valuation from $650 million to more than $2 billion in about a year. Its NodeZero platform autonomously runs penetration tests against production environments and now claims 120 percent year-over-year revenue growth, more than 310,000 safely executed production tests, and 7,000-plus protected organizations including four Fortune 10 firms, with FedRAMP High authorization. The round funds channel expansion, international growth, and autonomous remediation agents.
▸ The MSP Angle
Can I offer continuous penetration testing without hiring a red team?
Autonomous pentesting turns a scarce, expensive human service into software you can run continuously and package as a recurring line, and the fact that this vendor is putting fresh capital straight into channel expansion tells you where it wants that revenue to come from. The value for your clients is not the annual audit, it is finding the exploitable path before an attacker does, on a schedule a human team could never sustain. If security services are your growth lane, continuous automated testing is a high-margin offer to scope now, while the tooling is racing to sign partners.
Read at Help Net Security ↗
TechCrunch · Aug 4
Anthropic signs a $10 billion compute deal for a hydro-powered data center in Norway
Anthropic agreed to a $10 billion, six-year compute contract with Volta, a startup founded in January by former Brookfield executives that recently raised $300 million at a $2.4 billion valuation. The deal secures 133 megawatts at a Norwegian data center running on hydroelectric power and Nvidia's newest Vera Rubin chips, delivered in two phases targeting the end of 2026 and the first quarter of 2027. It follows Anthropic's separate multi-billion-dollar AMD partnership in July.
▸ The MSP Angle
Why should I care how AI companies buy their computing power?
Because the price and availability of the models you build client services on trace directly back to deals like this, and a lab locking up six years of capacity is a signal that compute is scarce enough to plan around. Scarcity and vendor concentration on the supply side become cost volatility and roadmap risk on your side. Price AI services on the outcome rather than the raw token cost, keep your platform able to route across model providers, and treat a single vendor's capacity crunch as a rerouting question rather than a client outage.
Read at TechCrunch ↗
PYMNTS · Aug 6
DeepSeek reopens an $8 billion raise and buys a stake in a humanoid-robot maker
DeepSeek resumed its second funding round on August 6, seeking close to $8 billion at roughly a $74 billion valuation, with Monolith Management reported in talks to contribute, after pausing the round over leaked founder remarks. Separately it took a 2.31 percent stake in humanoid-robot maker Unitree for 140.8 million yuan, about $20.8 million, through Unitree's Shanghai public offering, with the two planning joint models pairing DeepSeek's language systems with Unitree's motion control.
▸ The MSP Angle
Are cheap open models like DeepSeek actually usable for client work?
For cost-sensitive and privacy-sensitive client workloads, low-cost open-weight models are a real option, and a raise this size is a sign the downward pressure on model prices is not slowing. The catch is that open does not mean governed: running a cheap model well still needs the isolation, logging, and data controls a client's compliance posture requires. The opportunity is to offer clients the cost advantage of open models inside a managed environment you control, so they get the savings without inheriting the risk of pointing staff at a raw model endpoint.
Read at PYMNTS ↗
Ipsos · Aug 3
Ipsos: 67% of US workers expect AI to make the working experience worse, not better
An Ipsos poll released August 3, of 1,533 US adults aged 18 to 75 fielded in mid-June with a margin of error of 2.7 points, found 67 percent believe AI will worsen the national worker experience by eliminating jobs and increasing productivity pressure, against 30 percent who expect improvement. A fairness gap stood out: 51 percent said business owners and executives will benefit only or mostly, and nearly half of college-educated workers expect AI to help their own job versus one in five of those with a high school education or less.
▸ The MSP Angle
Why do AI rollouts stall even when the tools work?
Because two-thirds of the people being asked to use the tools expect them to make work worse, and a rollout that ignores that skepticism dies in the adoption phase no matter how good the software is. This is the change-management problem hiding inside every AI project, and it is exactly where an outside partner earns trust that an internal mandate cannot. Frame AI to client staff as capacity that removes the tedious work rather than headcount that removes them, involve the skeptics early, and measure adoption, because a licensed tool nobody opens is a cost, not a service.
Read at Ipsos ↗
CNBC · Aug 3
The White House AI review covers closed frontier models, leaving open-weight AI outside it
The White House convened AI companies this week to review its completed voluntary framework, mandated by the June 2 executive order and due August 1, which asks developers of covered closed-source frontier models to give the government up to 30 days of pre-release access for a classified assessment of the models' cyber capabilities. Open-weight models, increasingly capable and largely coming out of China, are cost-efficient and downloadable to run on any infrastructure, and sit outside the closed-source focus of the review.
▸ The MSP Angle
Does the new AI framework restrict the open models I might run for clients?
The review is aimed at the big closed frontier labs, so the open-weight models you might self-host for a cost-sensitive client are outside its pre-release gate for now, which is a practical advantage worth understanding. The tradeoff is that open models carry no vendor safety review, so the governance work moves onto you. Know which side of that line each client workload sits on: a closed frontier model comes with a vendor and a review, an open model comes with your controls and no one else's, and pricing the second one honestly means pricing in the oversight it requires.
Read at CNBC ↗
Gartner · Jul 27
Gartner: worldwide IT spending will hit $6.37 trillion in 2026, up 14% on AI demand
Gartner raised its 2026 worldwide IT-spending forecast to $6.37 trillion, up 14.2 percent year over year, crediting the increase to AI demand. The upgrade continues a run of forecast raises tied to AI infrastructure and services buildout across the market.
▸ The MSP Angle
Is there real budget behind all the AI talk, or is it hype?
A double-digit rise in global IT spending that Gartner attributes to AI is the macro number to put in front of a hesitant client: the budget is real and it is moving. That does not mean every client has a line for it yet, which is the opening. Bring the spending trend into planning conversations as evidence that peers are funding this, then convert the abstract tailwind into a specific, scoped AI service the client can buy this quarter rather than a someday intention.
Read at Gartner ↗
Anthropic · Aug 7
Anthropic retunes a safety filter to stop over-blocking legitimate biology research
On August 7 Anthropic refined the biology-safety classifier on its Fable 5 model, cutting false refusals, where the model over-blocks benign research questions, by roughly 85 percent while keeping its safeguards against genuinely dangerous requests in place. The change broadens usable access for legitimate research and clinical work.
▸ The MSP Angle
Why does AI sometimes refuse normal work questions, and is that improving?
Over-blocking is a real adoption killer for clients in healthcare, labs, and life sciences, where a model that refuses routine clinical or research questions trains users to stop trusting it within a week. This retune is a sign the vendors are treating false refusals as a bug worth fixing, not an acceptable cost of safety. When you deploy AI into a regulated or technical vertical, test the refusal behavior on the client's actual daily questions before rollout, because the model that blocks legitimate work quietly loses the room.
Read at Anthropic ↗
ChannelE2E · Aug 4
A regulated-vertical MSP launches an AI-native platform and a managed AI governance service
RFA, a managed services provider focused on financial services, launched an AI-native platform on August 4 that embeds AI-driven automation, monitoring, and analytics across cybersecurity, network, and IT operations for regulated clients, using AI agents to gather diagnostics, correlate signals, and clear repetitive ticket-queue work. Alongside it, RFA introduced a managed AI services offering to help firms deploy and govern the AI tools their staff already use, spanning strategy, implementation, governance, cost controls, and custom agents, with humans kept in the loop for security and compliance decisions.
▸ The MSP Angle
What does an AI-native MSP actually sell that a traditional one does not?
Two things, and this launch shows both: AI woven into the delivery model so the ticket queue partly runs itself, and a separate managed service that governs the AI a client's staff are already using. The second is the wedge most MSPs are missing, because clients do not need another AI tool, they need someone to put strategy, cost controls, and oversight around the tools they already have. A regulated-vertical firm productizing that is a signal the governance service is becoming a category, and the MSPs that name it and price it first will own it in their niche.
Read at ChannelE2E ↗
