InsideLLM Output Handling Needs No New Controls
Insecure output handling mitigations, layer by layer. Every LLM attack in this track dies to a control invented before LLMs, and the sandbox moves least.
[01] Recon — who you have found
The offensive security notes
I wish someone had left me.
I'm Bikramjeet Singh, a Lead Penetration Tester. Every engagement teaches something the courses skip over, so I write those bits down in full, web, API, Active Directory, cloud, Kubernetes and AI, including the attempts that went nowhere. No theory, no vendor filler, no "it depends".
[02] Enumerate — map the attack surface
Eight tracks. Four are live, four are next.
Everything I publish lands on one of these. The count is the honest depth, not a target.
S0142
S0221
S0306
S0401
S05
S06
S07
S08
Kubernetes Journey
Cluster security learned in public, start to finish. The series is complete.
AI / LLMs
Attacking and securing AI systems and the plumbing around them.
Threats & Exploits
Fresh CVEs pulled apart to root cause, not "patch now".
Web & API
App and API testing, auth flaws and business logic abuse.
AD & Infrastructure
Active Directory, lateral movement, privilege paths.
Coming soonCloud
IAM, storage exposure and config reviews that matter.
Coming soonSystem Breakouts
Kiosk escapes, restricted shells, thick client, mobile.
Coming soonWireless
WPA handshakes, evil twin, deauth, Bluetooth and RFID.
Coming soon [03] Exploit — the write-ups
What I learned recently
Newest first. Each one starts with something I did not understand, works through what actually happens under the hood, and ends with what it means for security. The wrong turns stay in.
LLM Hallucinations Are a Supply Chain Problem
A short reference on LLM hallucinations: the three types, the measured rates by model, and the checks to run before you install a package a model suggested.
PaperCut's Auth Check Ran, Passed, and Was Reading the Wrong Page
CVE-2026-81578 and CVE-2026-82078 chain to pre-auth RCE in PaperCut NG and MF, exploited as a zero-day. Why the auth check ran, passed, and still missed.
LLM Data Exfiltration: No JavaScript, Just an Image
LLM data exfiltration needs no script and no click. A markdown image is a silent outbound GET, and my own password reached my server in the query string.
[04] Escalate — the receipts
None of this is theoretical
Every line below is a link to something already on this site. No roadmap, no coming soon, just what is published and how much of it there is.
On the shelf right now
E01Kubernetes Journey 42 topics
E02A complete series: a cluster built, broken and rebuilt across 42 topics, from RBAC and admission control through to runtime.
AI / LLMs Journey 21 topics
E03How the models actually work, then how they fail, from neural network internals to attacking LLM apps.
Pentest Notes 351 notes
E04Command-level operator playbooks: 185 AI / LLMs, 140 Kubernetes, 26 Web & API. The bits you reach for mid-engagement.
Threat briefs 6 briefs
CVE teardowns taken to root cause rather than "patch now and hope".
Verified
OSCPOffSec
PNPTTCM Security
CASAApp Security
CKSKodeKloud course
Findings mapped to
OWASP OTGMITRE ATT&CKCWESANS Top 25CIS Benchmarks
Counted honestly
70write-ups
~15hreading
351notes
4live tracks
Started the Kubernetes track in the open and kept going for 42 topics. The counts move because I publish, not because I planned a number.
[05] Report — hand it over
If it was useful, tell me. If I got it wrong, tell me faster.
New write-up most days. Labs, dead ends and the odd working exploit. Compare notes, argue with me, or just come along for the ride.