Security Posturefix-this-first · ranked by leverage

Security · Posture

How safe the fleet is — and exactly what to fix first to move the score.

The board read on agent safety: one score, the single fix with the most leverage, the trend toward target, and the priced cost of the instrumentation gap.

Bottom line

Fleet posture 76/100 (target 80). Highest-leverage fix: “No tool-approval policy” on 5 agents = +50. ~2 weeks to target at recent slope. Board rec: instrument 1 shadow agent (worst-case ~69) and clear “No tool-approval policy” for the biggest single move.

Do nextFix “No tool-approval policy” on 5 agents (+50)Instrument 1 shadow agent →Harden worst agent — marketing-gpt →

fleet posture score

target ≥ 80

+50

top fix leverage

No tool-approval policy

weeks to target

at recent slope

~69

worst-case score

1 shadow → Critical

Fleet posture score

instrumented agents · target ≥ 80

76/ 100

0target ≥ 80100

Hardened

At risk

Critical

Posture trend

last 6 weeks · target ≥ 80

w1w2w3w4w5w6

Up +18 over 6 weeks (58 → 76). At the recent slope, ~2 more weeks to hit 80.

Fix this first — ranked by leverage

ROI = agents affected × points each

Fix “No tool-approval policy” on 5 agents = +50 posture — most leverage available.

1.No tool-approval policy · 5 agents × −10

+50

2.Running as root · 2 agents × −25

+50

3.Approvals & sandbox bypassed (yolo) · 1 agent × −35

+35

4.Sandbox disabled · 1 agent × −20

+20

5.Permission checks bypassed · 1 agent × −15

+15

Effective posture — pricing the instrumentation gap

shadow agents have no telemetry · their posture is assumed

reported (instrumented only)

76/ 100

1 shadow agent excluded from this number.

→

worst case (shadows = Critical)

~69/ 100−7

If unmonitored agents are as bad as our worst, the fleet drops here.

Instrumenting Zed Agent closes a blind spot worth up to 7 points of uncertainty — you can't claim 76 while 1 agent is dark.

Posture regression — current vs worst observed

agents that have been worse than they are now

support-copilotworst observed 50 (Running as root)current 65↑ 15

Critical & at-risk agents

score < 80 · lowest first

marketing-gptApprovals & sandbox bypassed (yolo) · Running as root · Permission checks bypassed25 Zed AgentNo tool-approval policyshadow60 support-copilotRunning as root · No tool-approval policy65 Codex CLISandbox disabled · No tool-approval policy70

At-rest installations — coverage

4/5 instrumented · 1 unmonitored

AdapterVersionExecutableConfigCoverage

claude-cli2.1.161/Users/mark/.local/bin/claudepresentinstrumented

claude-desktop1.9255.2/Applications/Claude.app/Contents/MacOS/Claudepresentinstrumented

cursor-cli3.5.17/Applications/Cursor.app/Contents/MacOS/Cursorpresentinstrumented

cursor-desktop3.5.17/Applications/Cursor.app/Contents/MacOS/Cursorpresentinstrumented

zed0.229.0/Applications/Zed.app/Contents/MacOS/zedpresentunmonitored

Derived: hardeningROI · postureIndexTrend · weeksToTarget · shadowAgents — from app/data/c16-fleet.ts (qcontrol /api shape). Score = 100 − sum of FACTOR_PTS deductions.