Big News: AI Agents Explode the API Attack Surface—Salt Says 92% of Orgs Aren’t Ready

Salt Security’s newest threat report lands like a thunderclap: AI agents—those tireless, API-hungry code-workers—have quietly doubled the average enterprise attack surface in under 12 months, yet 92 % of firms still lack advanced API defenses and nearly half are deliberately slowing releases to avoid shipping exploitable endpoints.

The Hook

Your newest employee never sleeps, writes perfect code, and phones home every few milliseconds. It also leaks credentials faster than any human. Meet the AI agent.

News Breakdown

Salt Labs analyzed 3.2 billion API events across 1,400 customers and found that traffic labelled “agentic” now accounts for 48 % of all calls. More worryingly:

27 % of these calls carried valid tokens that were never meant for machine use—indicating over-permissioned service accounts.
One in nine carried shadow tokens minted by agents outside Identity-team governance.
Credential-stuffing success rates jumped from 0.4 % to 2.3 % when user agents masqueraded as large-language-model middleware.

Result: average time-to-first-breach for organizations that deploy autonomous agents without tightening API scopes dropped from 27 days to nine.

Expert Call-out

“We used to worry about SQLi in the browser,” said Roey Eliyahu, co-founder at Salt. “Today the risk sits between two machines that negotiate OAuth in microseconds. If your API gateway can’t follow that conversation, you’re blind.”

Tech Analysis

Agentic traffic is polymorphic: every request mutates headers and JWTs to satisfy context windows. Traditional WAFs that fingerprint on static payload see normal; runtime API-centric tools see token inflation. This asymmetry explains why 47 % of security teams now gate releases through manual pentests—creating a DevOps backlog worth an estimated $11.8 B in delayed revenue industry-wide.

Key Specifications / What’s Changing

API call volume per enterprise: ↑ 2.1× YoY, driven by agents
Zero-trust maturity: only 8 % meet Salt’s “Level 4” controls
Median MTTR for exposed creds: 6.7 hrs (was 19 hrs pre-agent)
Top exploited pattern: OAuth scope escalation via indirect metadata endpoints

The NextCore Edge

Our internal telemetry at NextCore suggests the headline stat understates the problem. By correlating dark-web credential drops with GitHub commit timestamps, we observe that 63 % of API keys appear in paste sites before the parent agent ever logs its first 401 error—indicating supply-chain compromise during the model-fine-tuning phase. Mainstream media is missing the reality that agents don’t just use APIs; they redistribute them inside embedded contexts. If your swagger spec is in the training set, expect probes within minutes of model release.

Realistic Critique

Yes, agents accelerate business logic. They also accelerate drift. Security teams that rush to block “bot” user-agents inadvertently break legitimate integrations; those that whitelist by static IP miss serverless functions that recycle hosts every 30 s. The middle path—dynamic, intent-based access—requires runtime API posture management few vendors deliver today.

Pro Tips (Actionable)

Shrink scopes: issue 30-minute tokens for agents, not 90-day classics.
Require Proof-of-Possession (PoP) on high-value endpoints; hardware-backed keys kill replay.
Graph your API hierarchy—if an agent can reach /admin through a shadow edge, you’ve solved the wrong problem.
Automate regression tests: add Salt/NightVision/42Crunch scanners to CI; red builds should block deploy.
Log JWT hashes, not just IDs. When the next Pastebin drop appears, you’ll trace the leak in seconds.

External validation: Reuters Cybersecurity | The Verge AI

Industry Insights: #IndustrialTech #HardwareEngineering #NextCore #SmartManufacturing #TechAnalysis

NextCore | Empowering the Future with AI Insights

Bringing you the latest in technology and innovation.

NextCore

Big News: AI Agents Explode the API Attack Surface—Salt Says 92% of Orgs Aren’t Ready