Daily News: March 24, 2026
Google unleashes Gemini AI on the dark web, Interloom raises $16.5M for tacit knowledge, Canvas launches AI teaching agent, NVIDIA partners on grid-responsive AI factories.
Google Threat Intelligence has launched a dark web monitoring service powered by Gemini AI agents, now in public preview. The system processes upward of 10 million posts daily, building organizational threat profiles and distilling relevant security risks with claimed 98% accuracy in internal tests.
- Scale: 8-10 million dark web posts processed daily
- Accuracy: Google claims 98% accuracy in threat identification
- Profile building: Gemini creates org profiles from public data, identifying VIPs, brands, tech stack
- Alert prioritization: AI tags threats by relevance—direct mentions vs. ambiguous connections
- 7-day history: Automatically generates alerts going back one week when first activated
The system works by having customers confirm their organization identity, after which Gemini builds a profile using publicly available information. It then continuously monitors dark web forums, marketplaces, and communication channels for relevant threats—initial access broker activity, data leaks, insider threats.
"We are now processing every post from the dark web using Gemini, and from there distilling down what threats actually matter." — Brandon Wood, Google Threat Intelligence Product Manager
Traditional dark-web monitoring tools rely on keyword matching and regex, generating 80-90% false positives. Google's approach uses AI to understand context—for example, recognizing when a criminal claims to sell access to "a large North American bank with 50,000 employees" and matching it to the customer's actual profile.
This is genuinely useful for security teams drowning in alerts. The false-positive reduction alone could save thousands of analyst hours. But there's an uncomfortable flip side: Google is now one of the few organizations with AI-powered visibility into the entire dark web. That's immense intelligence power. And the company admits that state-sponsored hackers from China, Iran, North Korea, and Russia are already using Gemini for cyberattacks. The tool cuts both ways.
Interloom Raises $16.5M to Solve AI's "Tacit Knowledge" Problem
Munich-based startup Interloom has raised $16.5 million in venture funding led by DN Capital to tackle what CEO Fabian Jakobi calls the "tacit knowledge" problem: most operational knowledge in enterprises is never written down.
- Funding: $16.5M seed round led by DN Capital, with Bek Ventures and Air Street Capital
- Problem: ~70% of operational decisions are undocumented—only in experts' heads
- Solution: "Context graph" built from support emails, tickets, call transcripts, work orders
- Customers: Commerzbank, Volkswagen, Zurich Insurance already live
- Competition: OpenAI, ServiceNow, Microsoft all building agent orchestration layers
Interloom's approach is to ingest millions of operational records and build a continuously updated map of how problems actually get solved within an organization—what it calls a "context graph." Think of it like Google Maps for business processes: just as Google learns optimal routes from traffic data, Interloom maps the paths that experts take.
"The most important person at the bank is the person who knows whether the documentation is right or not. They're often the lowest paid. But they determine quality." — Fabian Jakobi, Interloom Founder and CEO
At Commerzbank, Interloom analyzed millions of customer support emails and found the gap between documented and actual operational knowledge shrank from roughly 50% to 5%. At Zurich Insurance, it won a company-wide AI competition—beating 2,000 other startups—for an underwriting use case.
The "Great Retirement" is creating a knowledge crisis: 10,000 Baby Boomers retire daily in the U.S., taking decades of institutional knowledge with them. Interloom's timing is propitious. But the competitive landscape is fierce—Microsoft, ServiceNow, OpenAI, and everyone else selling AI agents is trying to solve the same context problem. Interloom's bet is that their context graph gives them an edge. It might—but they're competing against companies with vastly more resources.
Canvas Launches IgniteAI Agent for Teaching
Instructure, the company behind Canvas LMS, has launched IgniteAI Agent—a new agentic AI tool designed to automate "low-value" faculty tasks like rubric generation and assignment feedback. The tool stops short of fully automated grading, positioning itself as an assistant rather than a replacement.
- Focus: Automating administrative tasks, not replacing teaching
- Capabilities: Rubric generation, assignment feedback, course design assistance
- Privacy-first: Open infrastructure, educational outcomes prioritized
- Concern: Some experts worry about "dead classrooms" where computers teach computers
- Timing: Competing with other AI education tools entering the market
The launch follows Instructure's February announcement of IgniteAI's broader platform, emphasizing agentic capabilities that can handle complex workflows at scale while maintaining human oversight.
Canvas controls the LMS market in higher education—ignoring AI wasn't an option. The "privacy-first, open infrastructure" positioning is smart, addressing institutional concerns about data lock-in and algorithmic transparency. But the real test will be whether faculty actually use it. Many AI teaching tools have failed because they create more work than they save, or produce output that requires extensive revision. If IgniteAI genuinely reduces administrative burden without degrading quality, it could become essential infrastructure.
NVIDIA and Emerald AI: AI Factories as Grid Assets
NVIDIA and Emerald AI are partnering with leading energy companies to develop "flexible AI factories" that connect to the grid faster and generate valuable AI tokens and intelligence. The initiative aims to unlock up to 100 GW of grid capacity for next-generation AI infrastructure.
- Partners: NVIDIA, Emerald AI, EPRI, National Grid, Nebius
- Goal: AI data centers that operate as flexible, grid-responsive assets
- Capacity: Up to 100 GW of potential grid capacity unlocked
- Demo: UK trial showed AI clusters can reduce power demand by 40% in response to grid signals
- Vision: AI infrastructure that balances compute demand with grid stability
The partnership builds on earlier demonstrations in the UK, where high-performance AI clusters reduced electricity demand in real time in response to grid signals. The model treats AI compute as flexible load—something that can be scheduled or throttled based on grid conditions, similar to how industrial facilities manage peak demand.
This addresses a growing constraint: AI compute is energy-hungry, and utilities are struggling to keep up. Treating AI data centers as "grid-responsive assets" rather than pure consumers is smart infrastructure thinking. The UK trial proved it works technically. The question is whether the economics pencil out—can AI companies make money by being flexible? If grid operators pay for demand response, maybe. But it requires AI companies to accept that sometimes, their compute will run slower or pause entirely.
What This Means
AI is getting better at finding what matters. Google's dark web monitoring processes 10 million posts daily and filters to relevant threats. That's not replacing security analysts—it's giving them signal instead of noise. The same pattern holds in other domains: context graphs for business processes, AI teaching assistants for administrative tasks.
Tacit knowledge is the new battleground. Interloom raised $16.5M to solve what everyone building AI agents is discovering: AI agents are useless without organization-specific context. The race is on to capture institutional knowledge before it walks out the door with retiring workers.
Infrastructure is catching up to AI demands. NVIDIA partnering with energy companies on grid-responsive AI factories isn't just PR—it's recognition that compute growth can't outpace energy infrastructure forever. The companies that figure out flexible demand will have a competitive advantage.