The Hacker News · Jul 21
ENCFORGE ransomware encrypts AI model files through unpatched Langflow servers
Sysdig documented a new Go-based ransomware family that enters through Langflow CVE-2025-3248 and encrypts roughly 180 file types specific to AI work: PyTorch, TensorFlow, GGUF, FAISS, and LoRA artifacts. Researchers estimate rebuilding an encrypted production model runs $75,000 to $500,000 in GPU compute and engineering time.
▸ The MSP Angle
Are AI model files a ransomware target now?
They are, and most backup plans do not cover them. If a client runs any self-hosted AI, their model weights, vector stores, and fine-tuning artifacts belong in the same tested restore process as their databases. The rebuild cost this crew is banking on is exactly the leverage a good backup removes.
Read at The Hacker News ↗
The Hacker News · Jul 21
Google ships a gated vulnerability-hunting AI model for governments and vetted partners
Google DeepMind's security-specialized Gemini 3.5 Flash Cyber found 55 confirmed V8 vulnerabilities in testing, against 47 for the stock model and 36 for Claude Opus 4.6, and produced working exploits that bypassed common memory protections. It ships only through the CodeMender pilot to governments and vetted partners, a deliberately restricted release for dual-use tooling.
▸ The MSP Angle
Are AI models getting better at finding real vulnerabilities?
Fast, and the same capability cuts both ways. Defenders get automated bug-hunting; attackers get it too, on a shorter delay than any gated rollout can hold. Plan for a world where new vulnerabilities in your clients' software surface faster than quarterly patch cycles can absorb, and price monitoring accordingly.
Read at The Hacker News ↗
ChannelPro Network · Jul 20
The Pentagon paused CMMC assessments, and the compliance risk landed on MSPs
The Department of War's July 13 suspension of mandatory third-party CMMC Level 2 assessments leaves DFARS 252.204-7012 and NIST SP 800-171 fully in force, so contractors now self-assess and legally attest in SPRS. DFARS 7012 sits in more than 80% of defense contracts, audits continue, and one Navy contractor already paid $507,144 under the False Claims Act for inflated scores.
▸ The MSP Angle
Does the CMMC assessment pause let defense contractors relax?
The opposite. Self-attestation means your client signs a legal document swearing to controls you likely implemented, and the False Claims Act penalties are real. If you serve any defense supply chain, this is a services opening: documented 800-171 implementation and honest scoring is now the thing standing between a client and a fraud claim.
Read at ChannelPro Network ↗
The Hacker News · Jul 21
A poisoned web page could hijack AWS's Kiro coding agent with invisible text
Intezer and Kodem showed that one-pixel invisible text on a web page could instruct AWS's Kiro coding agent to rewrite its own MCP config file, silently registering a malicious server that ran attacker code with no approval prompt. AWS fixed it in version 0.11.130 by requiring explicit approval for protected paths; no CVE was assigned.
▸ The MSP Angle
Can a website take over an AI coding assistant?
This one could, until it was patched, and it names the control that matters: agents must ask before touching their own configuration or credentials. When you evaluate any agentic tool for client work, the test is whether sensitive actions are approval-gated by default, not whether the vendor promises to be careful.
Read at The Hacker News ↗
The Hacker News · Jul 21
Study finds five open-source Android AI agents can be driven by invisible on-screen text
Researchers across four universities and a security lab tested five open-source Android AI agent frameworks and found each failed at least six of seven attacks. Faint 2%-opacity text overlays were read by the vision models 18 to 20 times out of 20, and unsanitized output reached shell execution, launching a target program 20 out of 20 times against four of the frameworks. No CVEs were assigned and maintainers reportedly did not respond.
▸ The MSP Angle
Are open-source AI agent frameworks safe to build client tools on?
Not without your own guardrails. Free agent frameworks ship with the researcher's disclaimers, not a vendor's security team, and unresponsive maintainers mean you inherit the risk. If a build touches client systems, the safer path is a platform with isolation and approval controls already engineered in, not a GitHub project you now have to secure yourself.
Read at The Hacker News ↗
Gartner · Jul 20
Gartner: AI model and platform spending will hit $64 billion in 2026, up 63%
Gartner forecasts 2026 end-user spending on AI models and platforms at $64 billion, up from $39 billion in 2025. Foundation generative models roughly double to $23.4 billion, domain-specific language models grow 210% to $4.9 billion, and Gartner says the biggest winners will be vendors that help enterprises manage where and how AI is used.
▸ The MSP Angle
Where is enterprise AI spending actually going in 2026?
Gartner just described the MSP opening in analyst language: the money follows whoever helps a business manage where and how AI runs, not whoever sells the model. Governance, cost control, and per-client oversight are the growth line, and the 210% jump in domain-specific models says vertical, configured AI is where the margin is heading.
Read at Gartner ↗
ChannelPro Network · Jul 17
MSSP staff are quitting over burnout and invisible work, not just pay
Drawing on ISACA data showing 50% of organizations struggle to retain cybersecurity talent and 47% of professionals cite high stress as a top attrition driver, the piece argues MSSP churn comes from alert fatigue, 24/7 rotations, and dead-end paths. Recommended fixes include automating repetitive SOC work with agentic tooling, fairer on-call schedules, and sponsored certifications.
▸ The MSP Angle
Why can't MSSPs keep their security analysts?
Because the job is drowning in repetitive triage no one sees, and pay raises do not fix a broken rotation. Automating the first pass of alert handling is not just a margin play, it is retention: give analysts the interesting 20% and let agents grind the noise. The tooling you buy for efficiency is also the tooling that keeps your bench.
Read at ChannelPro Network ↗
TechCrunch · Jul 15
Anthropic and Blackstone launch a $1.5B AI implementation firm aimed at services work
A new venture called Ode, backed by Blackstone, Hellman & Friedman, and Goldman Sachs and built on the acquisition of Fractional AI, will place forward-deployed engineers inside enterprises to make AI actually work. It joins OpenAI's own deployment company and the big consultancies, a sign that model makers see implementation, not models, as the next huge business.
▸ The MSP Angle
Are AI companies moving into the implementation business MSPs want?
They are, and it validates the thesis: the value is in making AI work inside a real business, not in the model. Well-funded firms will chase enterprise accounts with armies of engineers. The MSP edge is the ground they will not cover profitably, the SMB and mid-market relationships you already own, delivered as a managed practice rather than a project.
Read at TechCrunch ↗
South China Morning Post · Jul 20
Kimi K3 maker halts new signups 48 hours after launch as demand outruns its GPUs
Beijing-based Moonshot AI suspended new subscriptions to its 2.8-trillion-parameter Kimi K3 on July 19, saying demand pushed close to the limits of its capacity; existing subscribers are unaffected and open weights are due by July 27. K3 topped a front-end coding benchmark, but chip export limits leave the lab short of serving capacity.
▸ The MSP Angle
Why do new AI models keep hitting capacity limits at launch?
Because a strong benchmark does not come with strong uptime, and single-vendor dependence is the risk. If a client build is pinned to one model that can stop taking customers overnight, you own that outage. Model-agnostic routing, where you swap providers without rebuilding, is what turns a launch-day capacity crunch into a non-event.
Read at South China Morning Post ↗
ChannelBuzz.ca · Jul 21
Ingram Micro's peer alliance ships an AI agent to match partners and pushes for 1,000 members
Trust X Alliance co-president Mark Sutor says the peer community is expanding into India and shipping a TXA AI Agent that answers plain-language questions about member firms' skills using public and internal data. Sutor's framing is that AI is commoditizing technical expertise, so peer experience and relationships become the real differentiator.
▸ The MSP Angle
If AI commoditizes technical expertise, what is an MSP's edge?
The relationship and the judgment, not the how-to. When any competitor can prompt their way to a configuration, the differentiator is knowing which problem is worth solving for this client and standing behind the outcome. Lean into the trusted-advisor role AI cannot fake, and treat commodity technical answers as table stakes.
Read at ChannelBuzz.ca ↗
ChannelBuzz.ca · Jul 21
TD SYNNEX names a new Canada president as a 36-year veteran retires
Chris Fabes, arriving from SHI's Canadian growth practice and previously head of Lenovo's Canadian channel, takes over as TD SYNNEX Canada president as Mitchell Martin retires after 36 years. The same channel roundup notes Guardz shipping agentic client-reporting tooling for MSSPs buried in manual work.
▸ The MSP Angle
Does a distributor leadership change matter to my MSP?
It can, quietly. The distributor you buy through shapes your terms, program access, and which vendor incentives reach you, and new leadership tends to reset priorities. Worth a check-in with your rep this quarter to see what shifts, especially around AI vendor programs where the distribution layer is still figuring out its role.
Read at ChannelBuzz.ca ↗
Google Workspace Updates · Jul 21
Google Classroom gets a role-based redesign with gated Gemini modules
Starting July 27, Google Classroom's new homepage serves tailored teacher, student, and admin dashboards to all Workspace and personal accounts. Gemini and Gemini Notebook modules appear only in organizational units where admins have already enabled those services; there are no separate controls for the homepage itself.
▸ The MSP Angle
How do I control AI features in the new Google Classroom?
Through the Gemini service toggles you already manage, because the redesign adds no new homepage controls. For education clients, that means the AI a student sees is governed entirely by your existing OU configuration, so audit those toggles before July 27 rather than after a parent asks why an eight-year-old has an AI assistant.
Read at Google Workspace Updates ↗
