−7 days

Google's estimate for mean time-to-exploit. Negative.

Mandiant's M-Trends 2026 puts the estimated mean time to exploit at −7 days: exploitation often starts before a patch exists. That is an average with a lot of caveats (below), but the lesson for AI keys is simpler and harder: a stolen key is used immediately. What you control is how fast you notice and how much one key can burn.

−7 days

Estimated mean time to exploit — Google Cloud / Mandiant, M-Trends 2026. Method and sample are not published in the blog post.

−1 day

Average time-to-exploit in Qualys TRU's separate analysis of CISA KEV remediation records (2022–2025). Different dataset, same direction.

0.74%

Of 48,172 vulnerabilities disclosed in 2025, 357 were remotely exploitable and weaponized (Qualys). Most CVEs never start a race.

~1h41m

How long LiteLLM's malicious 1.107.0 sat on PyPI on Oct 21, 2026 (11:39 → quarantined by 13:20 UTC, per Wiz).

Read the number before you repeat it

Negative time-to-exploit is real enough to plan around and easy to overstate. What it does and does not say:

The ledger — sourced incidents

Each entry carries its source link. Where a figure is an inference rather than a stated fact, the entry says so. Updated as new incidents verify.

What can one stolen key burn?

Your window is how long until you notice, plus how long until the key is dead at the provider. The burn is that window times what the key can spend per hour. All inputs are yours; the arithmetic is printed. Nothing leaves your browser.

Hours until you would notice misuse

Alert, bill, canary, or a vendor telling you. Billing emails often arrive after the spend.

Hours from noticing to key revoked at the provider

Include finding who owns the key and every place it is used.

Most the key could spend per hour ($)

Think rate limit × price. No idea? Test with your highest-tier model.

Hard cap on this key's spend ($, 0 = none)

Only a cap the provider or a proxy enforces counts.

Copies of keys exist in logs, chat or tickets

Keys are saved in a hosted tool's console or workspace

The same key is shared across tools, environments or teammates

Shrink the window — the LLMJACK family

You cannot patch faster than negative seven days. You can decide what a stolen key is worth and how fast you find out.

SHIELD

Finds leaked keys in your repos, logs and configs — and rates what each one can burn. Free scan, local-only. Quick one-time report? KEY/LEAK does that too.

Open SHIELD