The research plan named seven desired headlines: the shocker, the awareness gap, witnessed incidents, ex-contractor access, change-blindness, what's at stake, and building-yesterday accountability. Six hold or reframe cleanly. One (building-yesterday) tilts the other way and needs the honest version. On top of the seven, the data handed us four structural findings, the foundations-versus-continuity split, the confidence-coverage gap, the prioritization barrier, and the budget contradiction, that are strong enough to carry sections of the report.
HOLDS — The data confirms the headline as written.
REFRAMED — The data supports the finding once the framing is adjusted.
CONTRADICTED — The data points the other way.
H1 · SHOCKER (LEAD)
Sophisticated cyber posture, 1990s-era physical posture
HOLDS
Headline intended
"Only X% of AI lab leaders have all five basics of high-security visitor management in place." Or: "Y% of organizations doing sensitive AI work still use paper visitor logs."
Data says (N=294)
8.2%
have all five high-security basics in place (badge, ID verification, government-ID capture, watchlist screening, and escort). 19.0% still sign visitors in on a paper logbook (Q2).
Disposition
The maturity-gap reveal lands hard. The "all five basics" cut is the cleanest shocker: fewer than 1 in 10 have the full high-security stack, despite protecting frontier IP. The paper-logbook stat (19%) is the backup. Lead with the basics number.
§02 Part A
H2 · AWARENESS GAP (THE GOLD)
Leaders who didn't know the option existed
HOLDS
Headline intended
"X% of AI and R&D leaders didn't know watchlist screening for visitors was an option." Positions Envoy as the educator, not just a vendor.
Data says (N=294)
12.6–15.6%
didn't know each capability was an option, across all seven advanced controls (Q4). For denied-party / watchlist cross-referencing specifically, 15% didn't know it was an option. 9.2% cite "didn't know it was an option" as a reason gaps stay open (Q2-a).
Disposition
This is the gold, as the plan predicted. The "didn't know watchlist screening was an option" framing is directly supported (15% on the denied-party item). The narrow, consistent band across all seven controls makes it durable. Stands as written.
§04 Part C
H3 · WITNESSED INCIDENTS
Leaders have personally seen strangers inside
HOLDS
Headline intended
"Y% of AI and R&D leaders have personally seen someone in their lab they didn't recognize in the past year." Time-bound, witnessed-by-the-leader stats are press gold.
Data says (N=294)
67.3%
personally saw someone they didn't recognize and weren't sure should be there within the past 6 months (Q8); 77.9% report at least one unauthorized-access incident in the past year (Q7); 38.8% have personally seen a visitor or contractor reach a restricted area (Q9).
Disposition
The strongest stat in the study. The plan asked for "past year"; the data is even tighter, two-thirds within 6 months. This is the lead the report opens on. Stands, with the tighter window.
§05 Witnessed reality
H4 · EX-CONTRACTOR ACCESS
The contractor with the still-valid badge
HOLDS
Headline intended
"Z% of terminated employees retain physical access for more than 24 hours, and a third of leaders don't think that's a problem." The Envoy article's still-valid-badge story, framed for AI labs.
Data says (N=294)
56.5%
are concerned about terminated employees retaining access, and 66.0% about contractors accessing spaces unsupervised (Q6), yet only 48.3% revoke terminated-employee access within 24 hours and only 35.4% govern contractors with the same system as employees (Q2).
Disposition
Holds on the concern-versus-control gap. Just under half close terminated access within 24 hours, so "more than half don't revoke within a day" is the defensible version. The "third don't think it's a problem" clause maps to the ~44% not highly concerned about terminated access. Frame to the revocation gap.
§07 Worry concentrates
H5 · CHANGE-BLINDNESS
Sites shifted, programs didn't
REFRAMED
Headline intended
"Contractor and vendor access surged at X% of AI labs in the past year. Only Y% updated their security program to match." The hybrid-work reframe per Shannon.
Data says
26.2%
of those whose sites changed updated their process for more unverified people in the building (Q5-a, n=271). 49.3% saw hybrid work increase, 43.5% more external visitors per week, 31.6% more contractor-heavy days (Q5).
Disposition
Holds, but the "only Y% adapted" stat is among the subset whose sites actually changed (n=271), not all 294. Frame as: "Among AI labs whose sites changed, only 26% updated their process for more unverified people in the building." Reframe to the change-experienced base.
§06 Change-blindness
H6 · WHAT'S AT STAKE
AI models top the list of IP leaders fear losing
HOLDS
Headline intended
"AI models and training data top the list of IP that AI leaders fear losing, ahead of proprietary research and product roadmaps." A ranking story for AI-specific press.
Data says (N=294)
18%
rank AI models and training data their #1 IP concern, ahead of client/customer data (17%), proprietary research (16%), and product roadmaps (11%) (Q1, top-choice share).
Disposition
Holds as written, AI models do top the list. The sharper, more honest framing is how close the field is: the top four cluster within seven points, so "no single asset can be ring-fenced" is the stronger story. Lead with the ranking, support with the crowded-field nuance.
§01 IP at stake
H7 · BUILDING-YESTERDAY
Can't tell you who was in their building yesterday
REFRAMED
Headline intended
"X% of AI lab leaders can't tell you who was in their building yesterday." Flagged as a strong social/share headline.
Data says (N=294)
44%
cannot confidently say they could tell you exactly who was in their building yesterday (34% disagree, 10% neither agree nor disagree). A 56% majority agree they could, 43% strongly (Q3).
Disposition
The clean "X% can't tell you" framing overstates it: the majority say they could, so "can't" is a minority. The honest, still-strong version: "Nearly half of leaders at frontier-IP organizations can't confidently say who was in their building yesterday." Reframe to "can't confidently say," and pair with the confidence-coverage gap (H10), since the self-rating may itself be optimistic.
§03 Part B
H8 · PRIORITIZATION
The blocker is prioritization, not cost
HOLDS
Headline available
The biggest barrier to better physical security isn't budget. It's that leaders have considered the controls and never prioritized them.
Data says (N=293)
58.0%
have considered the missing controls but haven't prioritized them (Q2-a), nearly 3x the share citing cost ("too expensive," 20.8%).
Disposition
The "so what" that ties the whole report together and reframes the conversation away from budget. Pairs with H2: the gap is part awareness, mostly prioritization. Strong enough to be the closing argument.
§04 Part C
H9 · FOUNDATIONS VS CONTINUITY
Foundations are common, identity continuity is not
HOLDS
Headline available
Programs are built on the basics; the connective layer that makes them auditable is missing.
Data says (N=294)
27.6%
maintain a single visitor record across all sites, and 35.4% govern contractors with the same system as employees, against 51.7% with badge access on sensitive doors (Q2).
Disposition
The structural spine of the Envoy thesis: roughly half have a badge reader, only about a quarter have connected identity across sites. This is the product-shaped gap. Strong support.
§02 Part A
H10 · CONFIDENCE-COVERAGE
Confidence outpaces coverage on every statement
HOLDS
Headline available
Leaders rate their programs more highly than the underlying controls justify.
Data says (N=294)
58.2%
agree their program would stop an unauthorized person (Q3), yet only 51.7% have badge access on sensitive doors and 47.3% escort visitors at all times (Q2). Net-agree sits above coverage on all five Part B statements.
Disposition
Consistent across all five statements (net-agree 52–59% vs lower Part A coverage). Programs are graded on intent, not on what's instrumented. This is also the caveat that makes H7's self-rating suspect. Stands.
§03 Part B
H11 · BUDGET CONTRADICTION
The investment imbalance leaders won't admit
REFRAMED
Headline available
AI labs underinvest in physical security relative to cyber, and most won't admit it.
Data says (N=294)
83.3%
say their current physical-cyber balance is right; only 11.9% admit underinvesting in physical (Q13), even though 67.3% just told us they saw a stranger inside in the past 6 months. 78.2% say leadership treats physical IP risk as a recognized priority (Q10).
Disposition
Leaders don't confess underinvestment, so "won't admit it" is the honest framing: the imbalance shows up in the gap between stated comfort (83.3% say balance is right) and witnessed reality, not in self-reported regret. Reframe to the contradiction.
§08 Budget perception