Linux vsock UAF
Controlled write + arbitrary read in vmw_vsock
A race between the loopback workqueue and socket teardown creates a use-after-free in kmalloc-96. Confirmed against Linux 6.12.79 with KASAN.
independent security research · lahore, pakistan
Vulnerability Researcher · AI Security Engineer · Agentic Systems Developer
I build autonomous systems that find real vulnerabilities. My agents read source code, form attack hypotheses, confirm them with sanitizers, and write working exploits — without human intervention.
➜ whoami
Haris Hussain // Shadow Hex
➜ cat profile.json
{
"focus": ["memory safety", "agentic security"],
"runtime": "100+ hours autonomous",
"confirmed": "7 findings / CVEs"
}[+] research pipeline ready
01 / profile
Self-taught in offensive security, AI engineering, and the places where complex systems quietly fail.
Haris Hussain is a self-taught security researcher and AI systems engineer from Lahore, Pakistan. He began in offensive security and evolved into building AI-driven autonomous research systems.
Instead of repeating manual workflows, he turns source code reading, call graph analysis, sanitizer confirmation, exploit writing, and coordinated disclosure into systems where agents handle every step.
“Manual research is slow. The future is autonomous agents that reason like an experienced researcher — at machine speed.”
Memory corruption, heap OOB, UAF, incomplete-fix detection, source auditing, binary analysis
Multi-agent architecture, autonomous research pipelines, self-hosted model deployment, training data pipelines
C/C++, Python, ASan / KASan / UBSan, GDB, strace, ltrace, ptrace, DynamoRIO, call graphs
Burp Suite, black-box testing, autonomous pentesting agents, web and API security
Docker, Linux (Arch, Ubuntu), OpenAI-compatible APIs, self-hosted LLM deployment
02 / systems
Two production-running autonomous systems. Built to reason, remember, verify, and ship work — not to produce scanner noise.
primary system · hxzn / xor-1
Give it a GitHub repository, source path, or binary. It returns a confirmed vulnerability with a sanitizer trace, minimized PoC, root cause, suggested patch, and full disclosure report.
Session-scoped facts, hypotheses, and decisions survive context trimming. The agent resumes exactly where it left off.
get_coverage_map(), get_unread_ranges(), and fingerprint checks prevent redundant reads and expose unread code.
active_hypothesis.json records the exact file:line, gap, next action, confidence, and status every iteration.
Loop detection, archive_component(), failure handling, and agent state tracking keep investigations moving past dead leads.
[2026-02-14 03:18:22] tool=find_callers
args={"function":"update_decompress_brush","path":"/sandbox/freerdp"}
result=4 callers · 2 unread ranges
[2026-02-14 03:19:01] hypothesis.status=confirmedsecondary system · black-box
A target URL, web application, or API endpoint is enough. Agents fetch, inspect, form attack hypotheses, choose their own tools and next actions, validate the result, and produce a complete report.
03 / findings
Confirmed vulnerabilities and published advisories, ordered by impact. Every card links to the complete technical record.
A race between the loopback workqueue and socket teardown creates a use-after-free in kmalloc-96. Confirmed against Linux 6.12.79 with KASAN.
doCdataSection() bypasses beforeHandler()/afterHandler(), leaving handler depth at zero and enabling XML_ParserFree() during an active callback. ASan confirmed the UAF.
A malicious RDP server can drive palette reads beyond the guaranteed stream region in update_decompress_brush(), producing a 23-byte OOB read for 32-bit color.
The parser validates payload_len after removing attacker-controlled PLEN bytes, then copies PLEN bytes without accounting for them. A crafted RTP packet overflows the decode buffer.
A stale loop variable is reused as a write offset in the Raspberry Pi decoder. Crafted padding_right data shifts writes beyond the row guard. Fix accepted within 24 hours.
Integer truncation can make lane_size zero, while initialization still writes two blocks per lane. ASan confirmed the write at monocypher.c:783.
Automatic beacon registration accepts crossOriginExposed data without checking the fenced-frame reporting policy. A child frame can cross a boundary the root operator did not grant.
04 / notes
Technical writeups covering the path from first signal to confirmed impact and coordinated disclosure.
05 / contact
For research collaborations, offensive security roles, or conversations about autonomous systems.