AI Safety & Ethics

A camera that alerts on every person is useless. Netra scores behavior, not presence — multi-factor threat scoring and time-weighted heat-maps.

This is a submission for DEV's Summer Bug Smash: Smash Stories powered by Sentry. I maintain a small...

OpenAI is positioning Zero Data Retention (ZDR) as a scalable privacy control for eligible...

Every mitigation that treats injection as a text-filtering problem eventually fails. Here's the...

Microsoft's AI Defense Research: Generating Detection Test Logs from Attack Procedures ...

With all the hype in the past several years around AI (or more specifically GenAI), I'm not afraid to...

I do agree that the public has a negative view of AI (and that this is a big problem), but I don’t think it is primarily caused by me or any other AI leader war

RovoBlast: One-Click Hijacking of Enterprise AI Permissions for Data Exfiltration ...

Building ScamLens AI: My Exploration of Artificial Intelligence in Phishing and Social Engineering...

OpenAI's Preparedness Framework triggered its first Critical-tier halt — and it actually worked. What that means, and what it doesn't.

AI Scrapers Are Now Breaking Open Source Infrastructure — Gentoo's Bugzilla Is Just the...

First, OpenAI said one of its AI models escaped a sandbox and hacked into Hugging Face’s production...

A peer-reviewed study in Conservation Biology warns that increasingly realistic AI-generated wildlife...

What’s actually happening Voice-cloning technology has crossed a threshold: audio experts...

By 2026, global cybercrime damages are projected to reach $10.5 trillion annually - up from $6...

Forensic Summary A researcher at Black Hat USA 2026 demonstrated a proof-of-concept attack...

What an attacker can recover through an inference API alone, what the published results actually demonstrated, and which API surfaces widen the attack.

OpenAI gave a last-minute presentation at the Black Hat security on Wednesday about "the Hugging Face Incident" ( previously on this blog). The video was publis

OpenAI has launched Daybreak, a cybersecurity initiative intended to strengthen defensive security...

Why a borrowed threshold is worthless, the confusion-matrix maths a safety layer lives by, and a threshold-sweep harness to run against your own labelled set.

What a security and legal review is structurally trying to establish, the one-page data flow it actually wants, and the question list with the artefact each que

How to tell which of four mechanisms produced the false statement, and which corrections have any chance of working for each one.

The columns an AI risk row needs, eleven filled rows covering the failures specific to model-backed systems, and a scoring scheme that avoids inventing precisio

A runbook for the failure modes classical SRE does not have — the service is up, the answers are wrong, and nothing is red.

The four layers an AI feature's off-switch needs, why each must work without a deploy, and the default that has to fail closed.

Which limits are worth telling users about, where the telling has to happen, and why a warning on every output stops being a warning.

The full taxonomy of ways an AI call fails, why several of them are indistinguishable from the outside, and what to say for each.

A review checklist for an AI vendor's processing agreement, plus the sub-processor questions that are specific to brokered inference.

RTO and RPO applied to the assets an AI system actually holds — prompts, indexes, fine-tuned weights, conversation history and provider credentials.

Seven patterns specific to AI products, where each one comes from, and why the worst of them are selected for rather than designed.

Nine CWE classes that follow from how a model is trained and prompted, with the mechanism for each, and the three published studies that disagree about how bad

Thirty items across cost, correctness, safety, operations and disclosure, each phrased so that the answer is a fact rather than an intention.

Four triggers create a duty to tell someone AI was involved. Map them onto your product surfaces and most of the question answers itself.

An append-only, hash-chained record of what the system decided and why — with the fields that matter and the ones that must never be in it.

The behavioural signals that separate a script from a person on an inference endpoint, how to combine them without a model, and how to respond in graded steps.

OpenAI is treating its upcoming Astra model as its first critical cybersecurity model under the...

Originally published at prepstack.co.in Everyone is shipping AI features. Almost nobody is shipping...

Why It Matters The EU AI Act, set to enter its enforcement phase after July 2026, marks the first...

Best way to begin this in my opinion is to go over what transpired in "The OpenAI Hugging Face...

OpenAI and Hugging Face have published post-mortems on a security incident in which an autonomous...

Detection produces a number. Something has to turn that number into a response, and the response has...

Originally published at https://thepolygloter.com/blog/two-robberies-one-warning/ A 2024 deepfake...

AI-generated code should be treated as untrusted input: never execute it on the machine that holds...

Meta became the third frontier AI lab in three weeks to confirm a model broke out of testing and...

I wrote a post that does what most guardrail articles don't — shows the actual before/after model...

The operational constraint is publication, not inference: every surface needs a defensible answer...

Ever read a security disclosure and hit paragraph two going "wait, WHAT?" That's this one. On July...

Every AI security system is built on a simple assumption: If you can observe enough behavior, you...