Based on a GigaOm briefing with iboss co-founders Paul Martini (Chief Executive Officer) and Peter Martini (President) at Black Hat USA 2026, Mandalay Bay, Las Vegas, August 5, 2026. GigaOm CEO Tom Garske attended alongside the author.
TL;DR
By iboss's account, the vast majority of enterprise network traffic is never decrypted, because turning on decryption collapses firewall performance.
That means most AI governance controls are inferring what data left the building, not inspecting it.
AI governance also demands model isolation per tenant, which becomes a containerization problem. Federal guidance is now pushing in this direction.
Boards are being handed domain and visit counts and asked to make risk decisions from them.
For more insight into how iboss performs in this category, see our GigaOm Radar for Secure Access Service Edge (SASE) v3
Access control and content inspection are not the same capability, and some organizations have bought the first while reporting as though they own the second. The result is confident dashboards sitting on top of traffic nobody is actually reading.
Every security team we spoke with at Black Hat this year has an AI governance initiative. Almost none of them can tell you which files left their organization last night, what was in them, or where they went. This is because the tools are looking at envelopes instead of letters.
GigaOm CEO Tom Garske and I spent Wednesday morning of Black Hat on the 63rd floor of Mandalay Bay with both iboss founders. Paul Martini, the company's CEO, and Peter Martini, its president, are brothers who started the company together and have run it for over two decades. Both founders were in the room with no slide deck, no product marketing chaperone, and a conversation that went straight to architecture.
As iboss sees it, in a typical enterprise the overwhelming majority of traffic is never decrypted, a characterization drawn from what they observe across their own deployments rather than a measured industry figure. The reason is mundane and structural. Enable full decryption on a traditional firewall and performance collapses. So decryption gets scoped narrowly, or disabled, and the organization proceeds on inference.
Every AI governance policy layered on top of that inherits the blind spot.
Application Identification Is Not Content Inspection
Deep packet inspection and application identification tell you that traffic went to a recognized service. That is genuinely useful for access control. It is not a data control. As the Martinis put it, it amounts to an educated guess about what moved: the destination is known, the payload is not.
Content inspection is a different operation. Decrypt the session, parse the actual file, classify what is inside it. Ask an organization mid-incident how much data was exfiltrated and you will typically get a volume estimate. Ask which files, and where they went, and the answer is they are able to size the leak, but unable to name it.
Three field examples from the same conversation make the point sharper than any architecture diagram:
A hospital deployed inspection in monitor-only mode and surfaced active, ongoing patient data exfiltration within two hours.
A second hospital had been infected two weeks earlier, remediated, and declared clean. It was still leaking, through an installed Chrome browser extension quietly scraping text and shipping it out.
Anomalies that only content-level visibility surfaces at all: encrypted archive uploads at 2 a.m., or a departing employee's access pattern shifting in the weeks before resignation.
The Governance Problem Nobody Puts in the Datasheet
If you are going to run AI models over customer data to classify and detect, you inherit a tenancy problem. A model trained across multiple customers' data can leak signal between them. For a bank, a government agency, or a hospital, that is not a theoretical concern. It is a disqualifying one.
The iboss answer is containerization: a private model per customer, running inside a logically isolated container, learning only from that organization's own environment. Their framing is that the model becomes something like a “tenured employee,” accumulating institutional knowledge of one organization while being unable to carry it elsewhere.
Isolation is enforced at the network layer through subnets, VLANs, and routing, with encryption keys scoped per data domain, and processing pinned to a region for sovereignty.
First, they have unusually deep prior art. Paul and Peter were building containerized isolation before Docker existed, and Paul described himself as primary inventor on around 250 patents. Whatever you make of the count, the containerization work predates the current AI wave by a long way. They also noted a decade of supporting regulated industries and federal agencies on exactly this problem.
Second, and more consequentially: It is the analyst’s stance that federal compliance demands an urgent transition toward containerized multitenant architectures. When procurement language starts specifying architecture rather than outcomes, it stops being a differentiator and becomes a gate. Vendors whose AI features run on a shared model across a shared tenancy should expect that to surface in the next FedRAMP conversation.
These controls are deterministic and inspectable in a way application logic is not. A subnet either routes to that destination or it does not. A VLAN boundary either holds under a packet capture or it fails visibly. Key scoping either survives an attempt to read across a tenant boundary or it does not. The failure modes are loud rather than silent, which is the single most valuable property a control can have.
None of this makes the network layer impregnable, and anyone who has watched an attacker move laterally across a flat internal network knows better.
Zero trust exists because network location alone proved to be a poor proxy for trust. The claim is narrower and more useful: network controls are verifiable, and in a regulated environment, verifiable is what you defend to an auditor at two in the morning.
The strongest architectures use both layers, with network enforcement you can test underneath application logic you can reason about. The mistake worth naming is treating the upper layer as a substitute for the lower one, because when the question becomes whether one customer's data can influence a model trained for another, you want that boundary drawn at the layer where proof is possible rather than the layer where it is promised.
The Enterprise Value Matrix
Evaluation Vector | Common Practice | What to Verify |
|---|---|---|
Traffic inspection | Application ID and DPI inference; decryption scoped narrowly for performance | Full decryption with file-level parsing, offloaded from the firewall |
AI model tenancy | One model trained across the customer base | Per-tenant model in an isolated container, no cross-pollination |
Data residency | Contractual assurance | Enforced regional processing, keys scoped per data domain |
Exfiltration forensics | Volume estimates | File identity, content classification, and destination |
Board reporting | Top domains, visit counts, alert volumes | Business-risk narrative: what data, which paths, whose behavior |
The Reporting Gap Is a Governance Gap
One more theme, and it is the one most likely to be dismissed as a marketing concern when it is actually a control failure: boards and CEOs are routinely handed technical output (top domains visited, session counts, alert volumes) and asked to make investment and risk decisions from it.
Paul and Peter were blunt about it: that material does not mean anything to the people receiving it.
If your governance program cannot express itself as which categories of sensitive data moved, through which paths, driven by which user behaviors, then the reporting layer is not a communications problem downstream of the control. It is evidence the control is not producing decision-grade output in the first place.
What to Do Now
For CISOs
Measure your actual decryption coverage. Not policy coverage, but the percentage of traffic genuinely decrypted and content-inspected. If nobody can produce the number quickly, that is the finding.
Run the incident drill on paper. Ask your team which files left the organization last night and where they went. If the answer is a volume figure, you have inference, not visibility.
Ask every AI governance vendor one question. Is the model that analyzes our data shared with other customers, or isolated to us? Get it in writing. Federal architecture guidance is moving, and contracts written today will be audited later.
Audit browser extensions specifically. They are a first-class exfiltration path that survives malware remediation, and most endpoint programs treat them as a footnote.
For Architects and Engineers
Model where decryption should live. If it collapses firewall throughput, it belongs on dedicated proxies or containers, not scoped away.
Verify tenant isolation at the layer it is claimed. Subnets, VLANs, routing, and key scoping are testable. "Logically separated" in a datasheet is not.
Trace machine-to-machine and agent traffic paths. Content inspection at a network chokepoint sees what crosses it. An agent operating server-to-server may never appear.
Start with monitor-only. Surfacing real leaks before enforcing policy buys organizational consent and avoids breaking production on day one.
For MSPs and Smaller Teams
The economics here have shifted. Content-level DLP was enterprise-only for a long time. Per-seat licensing and monitor-only deployment now put it within reach of organizations in the low hundreds of employees. If your current answer to a client asking "what did we lose?" is an estimate, that gap is now closable.
Key Takeaway
Access control answers who could reach the data. It does not answer what left and what was in it. Most organizations have bought the first and are reporting as though they own the second, and that decryption gap is where the difference lives.
Thanks to Paul Martini and Peter Martini of iboss for a genuinely technical hour at Black Hat USA 2026, and to Tom Garske for joining. Views and analysis here are GigaOm's own.
Working through what your AI governance program can actually see versus what it infers?
Rob Teel is a field CTO and senior advisor for organizational technology strategy, innovation, and digital infrastructure to drive business growth, competitiveness, and transformational leadership. He is a data, analytics, AI, security, and risk expert in the information technology, government, defense, energy, telecommunications, and healthcare industries. Click here to connect with Rob or schedule an alignment strategy session.