Manish Pandey Bengaluru
Grounded in classical security craft, I run adversarial tests and build threat models to secure AI and agentic systems. Read the math, not the press release.
Thirteen years on the security side of the house. Application code first, then infrastructure, cloud, the boring-but-load-bearing operational work, then, like everyone else who didn't see it coming, AI. The Data Science M.Tech from BITS Pilani is what lets me read the ML math behind the security headlines. The rest came from doing the job in places where the cost of a wrong call is real.
ls ./practice
Eight areas built over 13+ years. They cluster, method, offense, program, infrastructure, and the writing reflects how.
MAESTRO-aligned reviews. Training pipeline integrity, model supply chain, inference-time risk. Data poisoning, model inversion, evasion, covered end to end.
STRIDE, PASTA, attack trees, plus the five-zone method I use for agentic AI. Picking apart how a system can fail before someone else does.
Manual and tool-assisted. Web, API, mobile. Findings tied to OWASP risks with reproducible POCs, not scanner output dumped into a report.
Source-level audits with reproducible POCs and remediation patches. Best paired with a threat model, so the review focuses where the model says risk lives.
Building the program, not running the scanner. Triage, ownership, SLA design, the boring metrics that move the needle.
Red-team and purple-team exercises. Includes adversarial testing of LLMs and agentic systems. The EchoLeak-class chains classical pentests don't catch.
AWS, Azure, GCP. IAM, segmentation, data, runtime, CI/CD. Findings calibrated to the actual workload, not a CIS checklist.
Pipeline integration. SAST, SCA, IaC tool selection. Security gates that don't break the build. Practical adoption sequences for teams of 10 or 1000.
tail -n 3 ./writing
Long-form notes from the field. Threat modeling, red teaming, the places where classical security thinking stops working.
Certificates do not tell you whether a system can be manipulated through its own inputs. The threat model, the model supply chain, retrieval, agents, and five incidents that should change your questions.
Read articleThe enterprise AI purchasing journey seen from the security gate: the five stages, the three evaluation layers, and the residency traps hiding in vendor documentation.
Read articleI designed a two-day workshop with AI red-teaming as a full module, not a closing talk. We finished three of eight sessions. The classics ate both days.
Read article