Cybersecurity
Heartbleed
Critical 2014 vulnerability in OpenSSL's TLS heartbeat extension that allowed remote attackers to read up to 64 KB of process memory per request, leaking private keys, session tokens, and passwords from a large fraction of the public web.
Shellshock (Bash vulnerability)
Family of 2014 vulnerabilities in GNU Bash allowing remote code execution by smuggling commands through specially crafted environment variables. The original bug had been latent in Bash for roughly 25 years before discovery.
AES: The Encryption Standard That Protects Almost Everything
AES is the symmetric block cipher standardized by NIST in 2001, operating on 128-bit blocks with 128/192/256-bit keys — used in disk encryption, TLS, and CSPRNGs.
OWASP LLM Top 10
The OWASP Top 10 for LLM Applications is a community-maintained list of the most critical security risks for systems that integrate large language models, modeled on the long-running OWASP Top 10 for web applications.
Monoculture Risk in Software Security
Dan Geer's 2003 paper 'CyberInsecurity: The Cost of Monopoly' argued that Microsoft's operating-system monopoly made the internet systemically fragile — single vulnerability, vast blast radius. Geer was fired from @stake the same day the paper was published. 22 years later, AI-generated personalized software may be the actual practical antidote: if every business has custom AI-built CRM, exploits become per-target rather than mass-attacks.
CSPRNGs: Why Cryptographically Secure Random Numbers Are Essential for Security
Cryptographically Secure Pseudo-Random Number Generators (CSPRNGs) produce output that is computationally indistinguishable from true randomness — meaning an attacker cannot predict future outputs even with knowledge of past outputs. Standard PRNGs (Math.random(), rand()) are deterministic and reversible. CSPRNGs use entropy from hardware sources (timing jitter, thermal noise) and cryptographic algorithms to produce unpredictable output. Always use CSPRNGs for tokens, keys, session IDs, and any security-sensitive value.
Prompt Injection in LLM Systems
Prompt injection is an attack class where untrusted content fed into an LLM's context contains instructions the model follows as if from the developer. Indirect variants plant instructions in web pages, documents, emails, RAG indices, or images so the attacker never speaks to the model directly. The problem is structurally hard because LLMs cannot reliably separate instruction tokens from data tokens, and defenses range from fragile prompt-engineering boundaries to dual-LLM architectures and capability constraints.
Privilege Escalation: When Attackers Climb the Access Ladder
Privilege escalation is a security attack where a user gains access beyond their authorization — vertical (user→admin) or horizontal (user→other user) — a fundamental vulnerability class.
Indirect Prompt Injection
Indirect prompt injection is the variant of prompt injection where the attacker never communicates with the model directly; instructions are planted in content the model later retrieves, such as web pages, emails, documents, or RAG sources.
Jailbreak (LLM)
An LLM jailbreak is a user-supplied prompt that bypasses the safety training of the model itself, getting it to produce content it was tuned to refuse. Distinct from prompt injection, which targets the host application.
Negative-Day Vulnerabilities
Negative-day vulnerabilities — a term coined by Sherri Davidoff — are bugs being actively exploited before the vendor knows they exist. In an AI-augmented threat landscape where models can autonomously find and chain zero-days, negative-day is becoming the dominant mode of compromise — defenders are responding to breaches of vulnerabilities they've never seen described.
CPUID HWMonitor Supply Chain Attack (April 2026)
For 6-19 hours on April 9-10 2026, the official CPUID download links for CPU-Z and HWMonitor were modified via an Apache mod_rewrite path-traversal bug (CVE-2024-38475) to serve a trojanized installer hosting STX RAT — a Chrome password stealer. ~150 victims per Kaspersky telemetry. Same actor as the earlier FileZilla campaign.
ISO 27001: The International Standard for Information Security Management
ISO 27001 is the international certification framework for information security management systems, requiring risk assessment, 93 security controls, and continuous improvement.
Software Supply Chain Attacks: How Dependencies Become Attack Vectors
Software supply chain attacks compromise a software package or build system so that every application depending on it is automatically infected. The attack surface is enormous: a typical web application has hundreds to thousands of transitive dependencies, each one a trust relationship. Notable examples: SolarWinds (2020, compromised build system infected 18,000 organizations), event-stream (2018, NPM package maintainer handed off to attacker), and LiteLLM (2026, GitHub account takeover affecting 97M monthly downloads).
I-XRAY: The Demo That Turned Wearable Recording Into Real-Time Identification
In October 2024 Harvard students AnhPhu Nguyen and Caine Ardayfio chained Ray-Ban Meta glasses livestreaming to Instagram, face detection, the PimEyes reverse face-search engine, and an LLM scraping the results — returning a stranger's name and details in seconds. Nothing in it was novel, which was the point: commodity assembly, not a breakthrough. Never released, unmaintained past late 2024, and published with database opt-out instructions.
STX RAT Malware
STX RAT is a Zig-compiled Windows RAT deployed via Inno Setup + malicious CRYPTBASE.dll sideloading, using NTDLL proxying and IPv6-encoded .NET deserialization. Primary objective: stealing Chrome saved passwords via the IElevation COM interface. Campaign traced to at least July 2025, used in FileZilla and CPUID HWMonitor supply chain attacks.
Salt Typhoon SD-WAN Attack: Zero-Day Plus Firmware Downgrade Chain by Chinese State APT
Cisco Talos disclosed in February 2026 that Chinese state-sponsored actor UAT-8616 (Salt Typhoon/RedMike) exploited a CVSS 10.0 zero-day in Cisco Catalyst SD-WAN (CVE-2026-20127) to join victim networks as a fake branch, performed a firmware downgrade to reintroduce a 2022 vulnerability (CVE-2022-20775) for root escalation, then re-upgraded the firmware to hide forensic evidence. The campaign ran for at least three years before detection.
LiteLLM Supply Chain Attack: 97 Million Monthly Downloads Compromised via GitHub Account Takeover
In early 2026, LiteLLM — a Python package downloaded 97 million times per month — was compromised when attackers took over the GitHub owner's account and pushed version 1.82.8 containing credential-stealing malware. The malicious code hijacked Python's .pth startup mechanism to exfiltrate SSH keys, cloud credentials, crypto wallets, and database passwords from every infected machine. The incident exposed multiple systemic failures: transitive dependency risk, AI-generated bot spam suppressing the vulnerability report, and meaningless compliance badges from a provider accused of faking reports.
Anti-Camera Tech Misconceptions: The Gap Between Fiction and What Exists
'Invisible to cameras' is never literal — every real tool defeats one capture mode. CV Dazzle no longer works against modern networks but is still circulated as current advice; most AI-invisibility clothing is brittle; photo cloaking cannot help with already-scraped images; and beating facial recognition isn't anonymity when gait, voice and device signals remain. Metamaterial cloaking is narrow-band lab work, thermal camouflage is much closer, and the eye-implant scrambler has no real analogue.
IR-Blocking Eyewear and Retroreflective Clothing: Attacking the Sensor Instead of the Model
Much biometric capture uses near-infrared — 3D face mapping, iris scanning, night CCTV illuminators. Reflectacles (Scott Urban, 2015) absorb most near-IR at 780–1400 nm so eyes vanish on IR cameras while looking like ordinary sunglasses; Japan's Privacy Visor instead uses IR LEDs to blind cameras. Retroreflective fabric returns flash straight to the lens, overexposing the wearer — but needs a light source co-located with the camera, so it does nothing against ambient-light video.
Image Cloaking and Data Poisoning: Fawkes and LowKey
Imperceptible pixel perturbations added before posting, so that any facial recognition model trained on your photos learns a wrong representation of your face — data poisoning at the training stage rather than an inference-time attack. Severe limits: no help for already-scraped images, requires near-complete coverage since one uncloaked photo undoes it, and defeated by retraining on more robust models. Best understood as raising the cost of building a model, not protection.
Adversarial Fashion: Clothing and Makeup That Attack Computer Vision
Garments and makeup exploiting adversarial examples — inputs crafted so neural networks misclassify them. CV Dazzle (Adam Harvey, 2010) worked against the cascade detectors of its era and is now dated; adversarial-patch knitwear and 'AI invisibility' clothing are brittle across models and angles; adversarial licence plates poison ALPR databases. The frontier is inconspicuous adversarial makeup. Two structural problems: shelf life under retraining, and the conspicuousness paradox.
Anti-Camera Technology: What Actually Works and What Is Still Fiction
Counter-surveillance splits into fooling the algorithm (adversarial patterns, photo cloaking — durable only until models retrain) and blinding the sensor (infrared eyewear, retroreflective fabric — durable but narrow, each defeating one capture mode). Metamaterial and thermal cloaking are real research and mostly lab-scale or vehicle-focused. Broadband invisibility and implanted eye scramblers remain fiction; the nearest real analogue is IR-blocking glasses.
Web Bot Auth: Cryptographic Identity for AI Agents
Web Bot Auth lets a bot prove its identity per request using HTTP Message Signatures (RFC 9421) with an Ed25519 key and a published key directory, replacing user-agent strings and IP allowlists — which are trivially forged and hard to maintain.
WormGPT and Dark-Web AI Hacking Tools
WormGPT is a dark-web jailbroken LLM sold to cybercriminals starting 2023 for Bitcoin — Sherri Davidoff's team at LMG licensed it (~$500, $50 early-adopter) and tested it against Magento e-commerce platform, presenting at RSA 2024. Progression: 2024-early 2025 tools needed expert tweaking; a year later autonomous exploit chaining is viable. The arms race between attacker AIs and defender AIs is real and accelerating.