Notes on AI prompt security.
How prompt injection actually works, why we chose the browser, what on-device inference costs you, and what the DPDP Act asks of AI tooling — written by the people building it.
Writing an AI usage policy employees will actually follow
Most AI usage policies are a page nobody reads, enforced by nothing. What changes when the policy is a small set of machine-enforced rules instead of a document.
CERT-In's six-hour rule: what actually counts as a reportable AI incident
CERT-In's directions were written before generative AI was a mainstream enterprise tool. Where a prompt-driven data leak fits its incident categories, and why erring toward reporting is the safer read.
The DPDP Act's ₹250 crore penalty, and what actually triggers it
The headline number from India's DPDP Act 2023 is a ceiling, not a fine — and it attaches to one specific failure. What the Schedule actually says, and how “reasonable” gets judged.
How Aadhaar and PAN detection actually works
Checksum validation catches what pattern-matching alone misses. A practitioner's look at how Aadhaar's Verhoeff digit and PAN's structure let a detector tell a real number from a random one.
Why we intercept AI prompts in the browser, not at the proxy
Network proxies, CASBs, and API gateways each solve a real problem. None can see an employee pasting client code into ChatGPT from a home laptop. Here's why.
Indirect prompt injection: the attack that doesn't come from the user
Direct prompt injection is typing "ignore previous instructions." Indirect injection hides the instruction inside pasted content — much harder to defend.
What on-device inference actually costs you
Running detection in the browser instead of a server has real privacy and latency advantages — and real limits on model size and accuracy. Both sides, honestly.
What the DPDP Act actually asks of your AI tooling
India's DPDP Act 2023 doesn't mention AI — that's exactly why it applies. A practical read of which obligations bite when employees paste data into a chatbot.
What the first week of shadow AI telemetry usually shows
Before you can govern AI usage you have to see it. What the first week of visibility data tends to reveal, and how to read it without overreacting.