See your first finding in under 30 minutes with a free authorized pentest. Start your free assessment
A scanner checks a list. It does not form a hypothesis, read a response it did not expect, or feed a thing it just found into the next probe. Enlight™ is the product we built that can do all three as shown in a recent run on a test bed that we spun up. The testbed was an OWASP Juice Shop container that we deployed onto the internet and pointed Enlight™ at.
The goal? Train our engine and show the capabilities of performing an automated assessment.
After giving Enlight™ a root target, we stepped back and waited for the results. Throughout the run, Enlight™ acts as a pentest operator that automates the majority of what pentesters would initially perform, and then more.
Unlike just pure LLM agent exploitation, the engine Ironwood built has a deterministic scaffold that allows transparency, audibility, and reproducibility in the assessment running.
In this current run on a vulnerable Juice Shop app, we show an example finding that revealed a critical risk and how Enlight™ got to that point:
/rest/user/login as worth an auth-bypass attempt → a hypothesis to test.' OR true-- → getting back a HTTP 200 and an unexpected token in the body.role=admin).As one example of how the engine performs a task in an automated assessment, against /rest/user/login, Enlight™ sent a SQL Injection request:
POST /rest/user/login HTTP/1.1
Content-Type: application/json
{"email":"' OR true--","password":"anything"}
The response returned a valid admin session JWT (role=admin, admin@juice-sh.op) and verified SQL injection was possible. Enlight™ emitted two linked findings, SQL Injection Auth Bypass (Critical) → SQLi Auth Bypass Yielded Session Artifact (High), promoting the JWT to a artifact for the additional reactions that will follow.
Like an actual pentester, credentials are saved and used to further get into a vulnerable system. That artifact flow, recon into enumeration into exploitation is the symbolic-reaction differentiator. In addition, it spiders across all paths to see if they are subject to a SQL injection then chains further on top of it.
The chain is reproducible and human-verified. This saves time for any operator performing an assessment by returning high graded artifacts and findings that can then be progressed further on in the assessment.
It just doesn't tell you that there is a SQL injection vulnerability, but uses that information to see what else an attacker can do with it.
64 findings were surfaced by Enlight™: 5 Critical, 14 High, 19 Medium, 8 Low, 18 Informational ones.
The High tier includes a login missing brute-force throttling, HTTP method override accepted, an exposed Swagger UI (/api-docs), cache deception, decompression-bomb handling, and content-type confusion findings among the group.
The Medium tier is an access-control and app hygeine risk map: seven file-upload endpoints, unauthenticated PUT/PATCH/DELETE, missing HSTS, forgot-password email enumeration, an MFA endpoint missing rate-limiting, DOM sinks without Trusted Types, JWTs missing an exp claim, and exposed Prometheus metrics.
Every item carries its request and response for teams to verify and not just trust blindly.
All of these findings surfaced during the assessment shows how a foothold becomes real damage.
Most of what Enlight™ found is doors left unlocked. Several endpoints accept create, change, and delete actions with no login required at all, so an outsider can alter or destroy records they should never be able to touch. The login page caps nothing, not password attempts, and not, on the two-factor step, verification-code attempts. The result is that an attacker can simply keep guessing until both give way. Another finding is that the session tokens the app hands out never expire, so a single stolen login works indefinitely instead of timing out.
Other findings provide a map for an attacker to build out their attack surface.
The application publishes its full API manual to anyone who visits (an exposed Swagger console), leaks its internal operational metrics, and, through the forgot-password flow, will quietly confirm which email addresses have accounts. That last one is how a precise target list for phishing or password-guessing gets built.
None of these is an exploit by itself; each one makes every other attack faster, cheaper, and harder to spot.
Enlight™ was built not just to hack systems but to be a force multiplier to experts and IT teams who need a better view of their attack surface so that they can adequately secure their assets.
It is the ordinary connective tissue of a real breach. Unlocked doors, free reconnaissance, and levers to disrupt all surfaced in a single unattended pass.
The obvious way to build an autonomous attacker in 2026 is to point a language model at a target and let it improvise. Benchmarks are candid about where that stalls: even the strongest LLM agents solve only a minority of exploitation tasks and “systematically fail at complex multi-stage attacks requiring hypothesis validation and cross-service coordination,” while producing findings that don’t exploit and running non-deterministically [1].
Given model advancements and news we see from the frontier labs, agents are definitely getting better, but we haven't seen any widespread use of autonomous attackers yet. Enlight™’s Catalyst engine is on its way to be hybrid neuro-symbolic: the neural layer keeps what models do well — proposing what to try and building exploits within guarded frameworks— and the symbolic layer adds what they lack: deterministic, reproducible chains; typed artifacts that carry state across every stage without a context window to forget; and the finger that actually pulls the trigger.
Explaining further on the nerual layer, Enlight™ generates reactions the more runs it does, so that it can be replicated across multiple target systems deterministically. What this means is that automated assessments are not constrained by the tokens or API calls of other AI companies. The engine itself, like LLM models, improves for everyone the more assessments it gets trained on. In addition it will start exploiting further into the chain, including post-access reactions, the privileged actions an operator takes once a credential is in hand, so a chain can continue into what the recovered identity unlocks and to use the credential to pwn the system further.
Framed against a human pentest, the engine’s advantages are the ones a tester is glad to hand off. Speed and breadth: a full multi-phase assessment, unattended, in one pass.
It also chains so artifacts flow stage to stage, so the output is an attack path, not a disconnected list.
A human still validates the edge-case detector classes, and still owns business context and the chained-impact narrative. The only difference is that Enlight™ does the mechanical majority — recon, probing, credential chaining, evidence capture — so that the experts can spend their hours on judgment. It's a tool to make operators sharper, not redundant.
Enlight™ is not a scanner that detects more, but an engine that maps an attack surface, reasons, forms a hypothesis, reads what comes back, recognizes a credential, and can chain it forward. All with the evidence to back every call.
Start an assessment at portal.ironwoodcyber.com.
[1] CyberExplorer: Benchmarking LLM Offensive Security Capabilities, 2026 — arxiv.org/abs/2602.08023