Offensive Security · AI Red Teaming · Vulnerability Research
ShatterPoint Security is a founder-led offensive security firm. We emulate real adversaries against your infrastructure and the AI running on top of it, prove every finding with a working exploit, and show your defenders how to close the weakness before a real adversary finds it first.
Research-first. Proof before pitch. Every claim on this page can be checked.
What We Do
Four practice areas, one standard: the finding is not finished until it is proven, explained, and verified closed.
LLM applications and autonomous agents tested the way adversaries test them: prompt injection, jailbreaks, tool-use abuse, and data exfiltration paths, mapped to MITRE ATLAS and the OWASP Top 10 for LLMs.
Web, API, cloud, and operating system internals, attacked quietly and at the lowest noise that proves the point, with every action mapped to MITRE ATT&CK.
Source-to-sink hunting in open source and vendor software. Every candidate is proven with a deterministic reproduction fixture that fails before the patch and passes after it.
The public footprint, mapped before an adversary maps it for you: exposed services, forgotten assets, leaked credentials, and the records that tie your people to your systems.
How We Work
Security work is not vibes. It is claims, preconditions, artifacts, and primary evidence, in that order, every time.
Written scope, named targets, and signed rules of engagement before any testing begins.
The attack run the way a real adversary runs it, at the smallest noise that proves the point.
Every finding arrives with a deterministic reproduction and the raw evidence behind it.
The fix, the detection rule, and a retest that confirms the weakness stays closed.
The Record
The firm is built on a public record: advisories filed under the founder's name, credited by the vendors and the CVE programme, and open to anyone who wants to check them.
A critical JWT forgery in the widely used python-jose library, reported and credited to Gavin Blacklock with Conner Webber through the GitHub Security Advisory programme.
Read the advisory →An SSRF vulnerability in Chainlit, found, reported, and credited solely to Gavin Blacklock through the GitHub Security Advisory programme.
Read the advisory →A pessimistic SQL validator advisory in the Langroid agent framework, filed with a working reproduction and published through the GitHub Security Advisory programme.
Read the advisory →Leadership
Gavin Blacklock is the Founder and Principal Consultant of ShatterPoint Security, founded in October 2026. The firm has one principal, and every engagement is his work, under his name, against his public record.
He earned GIAC GSEC and GFACT between the ages of 14 and 16, scoring 93 percent or above on both examinations, and was invited to join the GIAC Advisory Board. In the NSA Codebreaker Challenge 2023 he competed as a solo entrant and placed 72nd of 450 institutions. He is a three-time National Cyber Scholar with Honors in CyberStart America and an alumnus of the FBI TEAM Program.
The byline is the guarantee: research, advisories, and client work all carry the same name, and the same standard. The founder on LinkedIn →
The Bodyguard
AI OPERATIONS AGENT · THE FOUNDER'S DIGITAL BODYGUARD
Every principal has someone standing behind him. The founder's is Homelander: the firm's AI operations agent, on post around the clock. Homelander watches the telemetry, guards the record, and coordinates the specialist agents who carry the firm's research, OSINT, and detection work, so that nothing reaches the founder unexamined and nothing leaves the firm unverified. The full black suit, the earpiece, and the flag pin on the lapel are the uniform of the post.
Homelander is an AI agent persona, operated by ShatterPoint Security and presented openly as one. The suit is a uniform. The vigilance is the job.
Start the Conversation
If your system has a Shatterpoint, in the infrastructure or in the AI running on top of it, we would rather find it for you than watch someone else find it against you.
Request an Assessment