Shadow AI used to be a slide about ChatGPT in the browser. It is now a product category the mid-market platform vendors are shipping into, not just talking about. Sophos put a stake in that ground today.
On Oct 13, Sophos announced general availability of Sophos AI Defense. The pitch is blunt: find the AI tools and agents employees are already using, including the ones security never approved, then enforce policy where those tools run. It ships through Sophos Fusion and pulls signals from endpoint, network, identity and browser into one control plane (Sophos).
The problem statement is the product. Sophos cites Gartner that 69% of cybersecurity leaders suspect or have evidence employees are using unsanctioned agents and AI tools, Cloud Security Alliance that 82% discovered shadow AI agents in the past year, and Komprise that 90% of IT leaders worry about shadow AI risk. IDC’s oft-repeated projection: 1.3 billion AI agents by 2028 (Sophos, citing those sources). You do not need to love survey math to see the buying motion. Boards are asking which AI is running and what data is going into it. Most CISOs still cannot answer.
Sophos’s answer is not another pure-play AI security console. It is an attach onto Fusion for shops already on Sophos EDR, XDR or MDR. Real-time guardrails inspect prompts and uploads. Activity feeds the investigation workflows teams already run, including Sophos MDR analysts (Sophos). That is the classic platform move: turn a new risk class into an add-on, not a rip-and-replace.
Why this matters for the category. Pure-play AI security startups sell discovery, prompt firewalls, model gateways and agent runtime controls as first-class products. Platforms like Sophos, Microsoft, CrowdStrike and Palo Alto are folding the same language into the suites enterprises already renew. The mid-market and MSP channel is where that fight gets decided first. Sophos’s release leans hard on MSPs and under-resourced teams who will not buy a fifth console for “AI.” If Fusion can show shadow AI and block a bad upload, the pure-play has to win on depth, not on awareness.
The threat framing is useful, and a little self-serving. Sophos X-Ops says it tracked an attacker using AI agents inside a customer network to build and test nearly 80 malware modules, plus underground trade in frontier-model API keys and jailbreaks (Sophos). That is a real attacker pattern. It is also the pitch deck for why AI Defense belongs next to XDR, not in a lab.
Where we push back. Visibility across four layers is necessary. It is not sufficient. Shadow AI is often a SaaS and identity problem: OAuth grants, browser extensions, corporate accounts on public models, and agents with tool access. Endpoint and network signals catch a lot. They miss the agent that never touches a managed laptop. Ask Sophos which sanctioned and unsanctioned SaaS AI apps it inventories out of the box, how prompt inspection works for browser-only tools, and whether findings land in your SIEM with enough context to write a ticket.
Pricing is Undisclosed in the launch note beyond “available as an add-on” for Fusion-enabled EDR/XDR/MDR customers. That is the real negotiation. AI security attach that shows up as a line item on the MDR renewal will win shelf space. AI security that needs a separate proof-of-concept will lose to whoever is already in the stack.
What we’d do this week. If you run Sophos, ask for the AI Defense SKU, the Fusion upgrade path, and a 30-day inventory of AI tools in a pilot OU before you buy anything else. If you are mid-evaluation with a pure-play AI gateway or DSPM vendor, force a side-by-side on shadow-AI discovery coverage, not on feature lists. And if you have no AI inventory at all, start there. The vendor can wait.
The read. The AI security market is splitting. Labs and regulated buyers will keep paying for depth. Everyone else will get “good enough” from the platform they already own. Sophos just told that second group the feature is shipping.