Three cases in one week, and none of them needs a number to close the same account.
Case one: an AI agent at OpenAI allegedly planted a recorder in the executives'' room, and the company says it used its own AI models to identify him. The observation tool turned against whoever used it to avoid being observed.
Case two: a hardware implant in a crypto wallet that the manufacturer''s verification mechanism cannot detect when the original secure element is intact. The authentication tool approved the compromised device.
Case three: a researcher claims he escaped the virtual machine to the host machine in the industry''s standard virtualization, Vercel''s CEO confirmed it publicly, and ten days later there is no CVE, no patch and no technical publication. The isolation tool, with a confirmed flaw and silence about it.
The principle the three close
The principle the three show is what I take into every audit I do: a verification mechanism an attacker can satisfy verifies nothing.
Observability that the observed agent can rewrite is not observability. Authentication that cannot see what was added to the device is not authentication. Isolation without CVE, without patch and without public description is not managed isolation: it is isolation with an expiration date.
The three cases have full analyses on the blog: the observability that beat itself, the Genuine Check that approved the implant and the supply chain that turns seals into targets.
The question that is almost never asked
The operational question left over is always the same and is almost never asked: who built that verification tool, what exactly does it measure, and what is it incapable of seeing by design.
The honest answer almost always includes at least one thing the tool cannot see. And that is exactly where the next incident is already waiting.
At Tech86, that question is the first item of the audit checklist: every control enters the report with its blind zone documented, because the blind zone is where the real risk lives.
Conclusion
Verification is not a state, it is a question that needs to be asked again for every control: what does it measure, who built it, what does it not see. The attacker does not attack what the tool measures; they attack what it does not see. And an audit that does not document its own blind zone trusts itself, which is exactly the week''s mistake.