This week the AI story stopped being about what the machines can do and became about who you should have trusted. Apple filed 41 pages accusing OpenAI of building its hardware on stolen secrets, and Meta quietly killed a three-day-old feature the moment Hollywood pushed back. Follow the thread; every story below is someone finding out, too late, who they let in.
The name Apple put in writing
Apple’s 41-page federal complaint calls OpenAI’s hardware program “rotten to its core” and puts a former Apple VP, Tang Tan, at the center of the allegation. A complaint is one side’s story, which is exactly why the specific claims are worth reading before you pick a side. Read the complaint →
Live for three days
If Apple is a fight over what was taken, Meta is a fight over what was consented to. Meta launched Muse Image’s @-mention feature and disabled it 72 hours later, after CAA and SAG-AFTRA leaned in, but the model itself is still running. What Meta pulled and what it quietly kept tells you which one it actually cared about. See what stayed →
The tools nobody approved
Those were outsiders. The harder problem is already inside the building: your employees are wiring personal AI accounts to company data faster than IT can even write them down. The instinct to ban it is the one move that reliably fails, and what works instead runs against every security reflex you have. Read why blocking backfires →
Sixty-seven percent unwatched
The shadow tools are run by people. The next blind spot is run by nothing at all. Orchid Security counted the nonhuman accounts inside enterprises and found 67 percent of them unmanaged, the precise gap AI agents are now moving into. The agents did not open the door; they inherited it. See the gap →
Two nines, one door
If the identities go unmanaged, the layer that guards them had better hold. This week it did not. Two CVSS 9.2 flaws let an unauthenticated attacker take over the BeyondTrust appliances that gatekeep everyone else’s privileged access, and that access layer is under sustained fire. When the thing you trust to fail last fails first, the whole chain is the story. Read the advisory →
One number
67 percent of nonhuman accounts sit unmanaged, and that single figure is the whole issue in miniature: the identities nobody watches are the ones agents and attackers reach first. Trust here is not a feeling, it is an inventory, and most of yours is uncounted. See what goes uncounted →
One thing to use
Shadow AI is the new shadow IT. Before you draft a policy, read the one explainer that treats the problem as inventory rather than prohibition; it is the practical field guide to what is already plugged into your data. Read the field guide →
That is the thread: stolen secrets, withdrawn consent, unapproved tools, unmanaged machines, and the access layer taking fire, all the same question about who got trusted and why. Hit reply and tell me which story deserves the full treatment next; I read everything. Forwarded this? Claim your own copy.
Dr. Joseph Joshua
P.S. If one story this week made you think of a colleague, forward it to them. That one gesture is how The Signal grows.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
