CVE-2026-33032 — an nginx-ui authentication bypass — was flagged on 480 instances in one customer environment; agentic pentesting could reach only 120 of them and confirmed exploitability on 80, leaving 360 instances without a direct verdict.
Agentic pentesting: speed and creative reach, with real limits
The rise of what the author calls Mythos-class models has compressed the window between a vulnerability disclosure and active exploitation to hours, and agentic pentesting emerged to match that tempo. In agentic pentesting, AI agents perform pentest steps dynamically and can discover exploit chains faster than a human tester, including chains a human might miss. The approach, as described by Sila Ozeren Hacioglu of Picus Security, trades deterministic repeatability for creative exploration: an agent may fail to exploit a vulnerability multiple times and succeed on a later attempt, runs are token‑intensive, and a single agentic campaign in a large organization can take weeks and cost materially more than deterministic scans.
Reachability and safety: where agentic runs stop
Agentic pentesting operates by running real attacks, so it is constrained to what it can safely reach. Safety rules preclude potentially disruptive exploits on fragile production systems, and physical or network segmentation leaves air‑gapped or isolated zones untested. Omdia research cited in the piece found organizations pentest only 32% of their attack surface on average, illustrating how much remains out of reach for active exploitation tests. The nginx-ui example makes this concrete: of 480 affected instances, agentic testing reached 120, confirmed 80 exploitable and 40 blocked, and left 360 instances—three out of every four affected hosts—untested because they were out of scope or unreachable.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleExploitability gaps beyond reach: missing PoCs and pre-public exploitation
Even on assets an agent can reach, not every CVE can be practically tested because many vulnerabilities lack a public exploit to run. VulnCheck data in the source found that only about a quarter of 2025 CVEs had proof‑of‑concept code or exploit details, and that 29% of the vulnerabilities VulnCheck confirmed as exploited in 2025 were exploited on or before their CVE publication dates. Asking an LLM to author an exploit is an option teams consider, but an untested exploit risks crashing services or corrupting data — precisely what safety rules aim to avoid.
Evidence‑backed exposure validation: how Picus Exposure Validation fills the blind spots
Where agentic tests cannot run or no exploit exists, the described alternative is to validate exploitability by testing the attacker behaviors a CVE requires against deployed controls and asset context rather than by attempting a live exploit. Picus Exposure Validation decomposes a CVE into attacker behaviors and checks whether those behaviors would succeed against live security controls within hours. In the nginx-ui example, that method produced verdicts for the 360 previously unreachable instances: 260 were assessed as blocked by controls and 100 as likely exploitable. Combined with the 80 confirmed exploitable hosts from agentic testing, the 480 raw findings became 180 evidence‑backed priorities—80 confirmed exploitable and 100 likely exploitable—so teams get an actionable, controls‑based triage rather than a partial scan result.
Discovery, validation, mobilization: a single workflow for faster response
The source lays out a three‑step cycle intended to avoid siloed or duplicated findings: discovery (enriching a CVE with threat intelligence, mapping attacker techniques, and identifying affected assets and controls), validation (Autonomous Pentesting, Breach and Attack Simulation, and Exposure Validation working from shared evidence), and mobilization (automatic deployment of low‑risk mitigations, assignment of remaining mitigations to owners, and revalidation). The piece says Picus Swarm’s agents, orchestrated by Numi AI, pick up a critical CVE as a signal and that the whole cycle can run in hours rather than weeks.
What this means for security teams, procurement leaders, and attackers
- Security teams: The practical implication is to combine deterministic testing for repeatability and scale with targeted agentic runs for creative discovery, and to add exploitability validation to cover unreachable hosts and vulnerabilities without public PoCs.
- Procurement and platform owners: Platforms that unify Autonomous Pentesting, Exposure Validation, and Breach and Attack Simulation against a shared, evidence‑backed findings model reduce duplicated work and produce a single prioritized backlog, according to the piece.
- Attackers: The author stresses that attackers using Mythos‑class models can now weaponize a CVE within hours, compressing defenders’ response time and raising the need for faster validation workflows.
The record in this piece is straightforward: agentic pentesting expands what can be discovered and exploited quickly, but it does not, by itself, validate every affected asset or every CVE. Behavior‑based exposure validation closes those blind spots by producing evidence‑based verdicts where live exploitation is unsafe or impossible. For teams wrestling with shrinking windows between disclosure and exploitation, the prescribed path is operational integration — deterministic tests for breadth, agentic runs for depth, and exposure validation to reach what active testing cannot. The remaining question the facts leave is whether organizations will adopt a single, evidence‑shared workflow widely enough to match the speed attackers now claim.
