Tech Graveyard/infrastructure
Static API Keys (1998 to Dying)
When one company runs 10,000 ephemeral AI agents, static API keys stop being a security gap and start being a security crisis.
Born 1998 · Still dying, projected end 2028 · Status: dying
By Deepak Gupta · Published · Last updated
Paired prediction: what replaces it in Future Tech
Certificate of Death
Name of decedent
Static API Keys
- Born
- 1998
- Died
- Not yet
- Age
- 28+
Cause of death
AI agent proliferation is making unscoped long-lived secrets impossible to govern
Survived by
SPIFFE/SPIRE, workload identity federation, short-lived JWTs, mTLS
Invented by
Early SOAP and REST API designers
The hook
GitHub detected more than 1 million leaked secrets in public repos in the first eight weeks of 2024, more than a dozen a minute. The fix is not better scanning. It is making the secret short-lived enough that the leak does not matter.
Thesis. Static API keys worked for 25 years because the blast radius was bounded. AI agents are breaking that boundary.
The story
The origin
REST APIs needed simple auth. The API key was the path of least resistance. Stripe, which launched publicly in 2011, popularized the model for developer-facing SaaS, and every developer-facing API copied the pattern: a long random string, scope-less, valid until manually revoked.
The peak
By 2018 every SaaS product had API key authentication and a 'keep this secret' warning in the docs that nobody followed. The keys ended up in committed code, CI environment variables, Slack messages, screenshots in support tickets.
The leak epidemic
GitGuardian counted 12.8 million new secrets exposed on public GitHub in 2023 and 23.8 million in 2024. Private repos add more. Its 2025 report found 70% of secrets leaked in 2022 were still valid. Breaches involving exposed credentials include Uber in 2022 and CircleCI in 2023.
The AI multiplier
When an enterprise deploys 10,000 AI agents, each needing access to multiple services, static keys become ungovernable. The volume crosses the threshold where individual key management collapses.
The replacement
Workload identity federation. Short-lived tokens issued per request. SPIFFE/SPIRE for cross-cluster identity. mTLS for service-to-service. The replacement is mature; what is missing is the migration tooling.
Key data points
- GitGuardian State of Secrets Sprawl 2025: 23.8M secrets leaked on public GitHub in 2024
- GitHub: 1M+ leaked secrets in public repos in the first eight weeks of 2024
- SPIFFE/SPIRE: CNCF graduated 2022
- Breaches involving exposed secrets: Uber 2022, CircleCI 2023
Contrarian angle
Secret scanning tools are a profitable industry built on top of a broken model. The right move is not better scanning. The right move is making secrets that cannot be useful when leaked.
The flip side
What replaces it
The paired prediction in Future Tech.
Read the prediction◆ Technology Rewired
See the full evolution in Technology Rewired:
FAQ
Are personal access tokens the same problem?
Yes. PATs are static API keys with a fancier UI. The same blast-radius issues apply. GitHub has started shipping fine-grained PATs with expiration, which is the right direction.
How do AI agents get identities if not through API keys?
Workload identity federation. The agent runtime presents a verifiable attestation (cloud IAM token, OIDC claim) and exchanges it for a short-lived credential scoped to the specific tool call.
Is OAuth client_secret a 'static API key'?
Functionally yes, when it is long-lived. FAPI 2.0 requires private_key_jwt or mutual TLS client authentication instead of shared secrets, for exactly this reason. Short token lifetimes do the rest.
More from guptadeepak.com
Further reading
Related obituaries
More from the infrastructure graveyard.
1999 to dying
ZombieOn-Prem Active Directory
Active Directory is the most successful piece of enterprise software that nobody chooses anymore. It is the primary directory at an estimated 90% of the Fortune 1000, and almost no new deployments.
Cause: Remote work and SaaS adoption made the on-prem domain controller indefensible
infrastructure · Peak 2015 · Projected end 2032
1996 to dying
DyingThe Corporate VPN
The VPN was the moat around the castle. Then everyone left the castle. Zero Trust networks do not have moats.
Cause: Remote work plus SaaS adoption made the perimeter model indefensible
infrastructure · Peak 2020 · Projected end 2028
1995 to 2022
DeadInternet Explorer
IE was the browser that defined the internet for half a generation. It died as the browser-as-OS model died. ActiveX, COM, and the whole intranet stack went with it.
Cause: Chromium-based browsers and the death of plugin-based web extensibility
infrastructure · Peak 2003 · Died 2022