Keynote by Adam Shostack: Stop Trying to "Manage Risk"

@burkhard wrote about this extensively in his latest newsletter.

Summary

Stop “Managing Risk.” Start Engineering Security.

The core idea

Cybersecurity “risk management” (likelihood × impact) usually can’t be done honestly, so we should stop pretending and treat security as an engineering problem with clear rules, thresholds, and outcome measurements instead.


The problem

  • Cyber incidents are one‑off and adversarial, so we rarely have good data for “likelihood × impact” estimates.
  • “Risk scores” become guesswork, fuel arguments, and still don’t reflect how leaders actually decide (cost, safety, uptime, regulation).
  • The word “risk” drags in an illusion of actuarial rigor that doesn’t fit our domain and slows down real security improvements.

What we should be doing

  • Change the language
    Talk about hazards, dangers, threats, and consequences. Reserve “risk” for the rare cases where you have actual measurement.
  • Use engineering rules, not gut‑feel risk
    Define explicit bug bars and release criteria:
    • “Any remotely exploitable vuln in a privileged process is release‑blocking.”
    • “Any vuln that can brick a device or endanger safety requires emergency response.”
  • Adopt shared thresholds
    Where possible, use (or push for) standard numbers instead of inventing “acceptable risk” locally every time, e.g.:
    • Target phishing click‑through rates.
    • Patch timelines by severity class.
    • Minimum MFA coverage ratios.
  • Measure outcomes like public health
    Track incident rates and how they move when you add controls, for example:
    • Credential‑phishing incidents per 10,000 users per quarter, before and after rolling out FIDO.
    • Remote‑code‑execution incidents per 1,000 devices per year, before and after hardening.