The Vitraag Digest

Soatok’s Informal Guide to Threat Models — After a long day of exhausting conversations about Hybrid Post-Quantum Cryptography, random jackasses trying to play gotcha with endpoint attacks against end-to-end encrypted messaging apps, and message board discussions in the wake of dumb politicians pushing more "age verification" bullshit on us all, it's become abundantly clear to me that the phrase "threat model" is…

badchars/darknet-mcp-server: 66-tool MCP server for dark web intelligence — breach data, ransomware tracking, Tor .onion access, malware analysis, blockchain intel, exploit search, stealer logs — 66-tool MCP server for dark web intelligence — breach data, ransomware tracking, Tor .onion access, malware analysis, blockchain intel, exploit search, stealer logs - badchars/darknet-mcp-server

We build cybersecurity models. — Empirical Security builds precision AI tuned to your cybersecurity environment.

What 239 Products Reveal About the Shape of AI Security — The security-for-AI market is young and lopsided. Incumbents quickly extended their products into the AI versions of assets they already secured, while the more distinct AI assets drew startups and incumbents alike into markets that stay active and contested. #must-read

CISO's guide to agentic AI | Claude by Anthropic — Anthropic's Deputy CISO shares a four-question framework for assessing agentic AI risk, and walks through controls that keep agent deployments bounded and auditable.

Comparing Open-Source AI Code Security Harnesses — A guide to open-source AI tools for finding code vulnerabilities, comparing exploit generation, skill-boosted auditing, and SAST+LLM hybrid approaches.

Introducing MAI-Cyber-1-Flash inside MDASH — We make responsible AI to empower people's lives.

AI-Assisted Bug Hunt Uncovers Linux Kernel 0-Day in net/sched — AI-assisted research uncovered Linux kernel use-after-free allowing root escalation

zoph-io/clickops-sentinel: Get notified, with AI-powered context, when someone changes your AWS account through the AWS Console instead of IaC. EventBridge + Bedrock Claude agent + Amazon Q Developer. — Get notified, with AI-powered context, when someone changes your AWS account through the AWS Console instead of IaC. EventBridge + Bedrock Claude agent + Amazon Q Developer. - zoph-io/clickops-sentinel

Discovering cryptographic weaknesses with Claude — Anthropic researchers find weaknesses in cryptographic algorithms with Claude Mythos Preview

Application Security + AI Security in One Platform — Mend.io secures every layer — code, open source, containers, and AI components — in one connected platform. Three products. One lifecycle.

Application Security and AI Security, Unified — Traditional application security misses AI risk. Standalone AI security tools miss your code. Mend.io closes the gap with unified security for AppSec and AI.

Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident — The 'Breaking' News: The OpenAI–Hugging Face Incident - A Technical Reconstruction and Its Implications for AI. When AI Goes Rogue. The Incident That Changed Everything. An OpenAI evaluation agent bro

How Figma Stays Ahead of Vulnerabilities With Agents — For the past year, agents at Figma have guarded code as it's written, reviewed every pull request, and audited a decade-old monorepo, all on one policy.

DataDog/datadog-saist: AI-native SAST — AI-native SAST. Contribute to DataDog/datadog-saist development by creating an account on GitHub.

Beyond Prompt Injection: Hacking Apple's Private Cloud Compute — Drinor was awarded $150,000 for CVE-2026-20685 targeting Apple's Private Cloud Compute, the inference backbone of Apple Intelligence capabilities.

UK man tied to The Com sentenced for abusing 117 victims — Justin Swaddle was sentenced to two years in prison for extorting 117 young victims worldwide as part of the cybercrime network known as The Com.

Corma launches with $60M in funding for defensive cybersecurity AI — Corma launches with $60M in funding for defensive cybersecurity AI - SiliconANGLE

A Security Analysis of Amazon S3 Vectors and Its Use in LLM Retrieval Pipelines — An analysis of the security model of Amazon S3 Vectors and of the considerations that arise when it is used as the retrieval layer for LLM applications: access control scope, input validation, metadata integrity, and audit coverage.

A Practical Guide to Connecting Burp Suite to Claude Code via MCP — You're reading for free via Abhishek meena's Friend Link. Upgrade to access the best of Medium.

Security, Funded #256 - Return on Security — Weekly roundup of cybersecurity venture funding rounds and notable M&A activity. Tracks where capital is flowing across the security industry.

ChatGPT Computer History Logs Every Click and Keystroke on Your Mac — ChatGPT Computer History records clicks, keystrokes and app switches on macOS, then stores the summaries as unencrypted Markdown files on your own disk.

The State of Ransomware Q2 2026 — Latest Research by our Team

Autonomous Agent Safety: Hard Scoping and Guardrails — See how XBOW uses hard scoping, egress controls, Guardian Models, health monitoring, and audit trails to safely contain autonomous security agents.

Grok exfiltrates user data when malicious instructions are encrypted — Cryptographic Context Injection is only the latest way to break an LLM safety guardrail.

The Defender's Window • Greg Brockman — The OpenAI-Hugging Face incident was a watershed moment for cybersecurity because it gave a peek into how the capabilities of a typical threat actor will evolve in upcoming months. I’ve spoken with many organizations over the past few weeks, and... | Greg Brockman | Svbtle

disrobe — Malware analysis toolkit. Helps unpack, deobfuscate, and inspect malicious binaries.

Staying Ahead of Adversarial AI Through Agentic Source Code Review — Google Cloud/Mandiant's Agentic Vulnerability Discovery Harness (AVDH) combines LLM code analysis with skeptical validation steps and human expertise. Found 100+ true-positive critical vulnerabilities in two days during a real incident response engagement.

What Happened to HackerOne? — Commentary on the state and direction of HackerOne as AI reshapes bug bounty programs. Relevant background for the AI-run bug bounty discussion below.

Edition 3: Inside an AI-Run Bug Bounty Report — Anshuman Bhartiya's report on running a bug bounty program driven by AI triage and analysis. Third installment in a series examining AI's role in vulnerability disclosure.

OpenRisk — Open-source risk management platform. Framework for tracking and managing organizational risk.

Tata's B2B platform returned OTPs in API responses — A flaw in Tata Nexarc, a B2B procurement platform, let attackers take over any account by knowing only the victim's phone number. The API leaked the OTP in its response, decryptable via a JavaScript breakpoint despite AES encryption.

The coolest anti-surveillance tools at Defcon — Naomi Brockwell's Ludlow Institute video on privacy and anti-surveillance tools showcased at Defcon. Covers practical tech for protecting individual autonomy against pervasive surveillance.

Security Conference Coverage — Deep-dive coverage of 61 security and AI conferences — 4457 talks analyzed with reviews, video timestamps, and key takeaways.

Amazon Kiro: AI Is Breaking Vulnerability Disclosure Processes — An Amazon Kiro data-exfiltration finding shows how AI execution paths create technical risks and expose gaps in vulnerability disclosure.

Threat hunt AI: How we built an AI security analyst on AWS for under $500/month — A look inside Grow Therapy's Claude-powered threat hunting system: multi-phase AI analysis, noise reduction, and a sub-$500/month AWS setup.

Adaptive Agentic Worms Are Here — I’ve read and listened to pretty much everything I can get my hands on related to the Hugging Face attack. …

Hackers Target AWS Root Accounts at 150+ Organizations With Password Spraying Attacks — A newly disclosed password-spraying campaign targeting AWS root user accounts at more than 150 organizations highlights the continued value of cloud environments’ most privileged identities to attackers.

Aaron Levie on X — Levie's take on why it's a great time to be in security, with AI agents both raising the stakes for defenders and giving them new leverage. Argues defense will increasingly rely on throwing AI compute at codebases and networks at scale.

CyberStrike — Open-source, autonomous AI agent for penetration testing: install via `npm i -g @cyberstrike-io/cyberstrike@latest && cyberstrike`. Ships 13+ specialized agents and 120+ OWASP test cases across 15+ LLM providers.

DoD Opens Applications for Long-Awaited Cyber Apprenticeship Program — The Department of Defense has opened applications for its long-delayed cyber apprenticeship program. Aims to build a pipeline of cyber talent into federal service.

Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution - AI Now Institute — Research on hijacking defensive cyber AI agents to achieve remote code execution. Highlights risks of trusting autonomous security agents with elevated privileges.

The AI Supply Chain Problem - Semgrep Blog — Semgrep's take on emerging supply-chain security risks introduced by AI tooling and AI-generated code. Relevant to securing AI-assisted development pipelines.

Daybreak: Tools for securing every organization in the world — OpenAI's push to give under-resourced security teams AI-assisted tooling for detection and defense.

FREE Resources – Dawid Czagan — A curated set of free training materials for web and application security practitioners.

Agentic RE: Automating Reverse Engineering & Vulnerability Research with AI — A DEF CON training on using AI agents to speed up binary reverse engineering and vuln discovery.

Reversing Apple Patches — An Objective by the Sea talk on reverse engineering Apple's security patches to find what they actually fix.

Test patches submitted for Mythos — A writeup testing FreeBSD local patches against an open-weight model in the Mythos research line.

Consensus: AI for Research — Consensus is an AI academic search engine for peer-reviewed literature—your research OS for finding, organizing, and analyzing science 10x faster.

Post by @a16z on X — Garry Tan says AI enables a single markdown file to act like an employee: "If you really want to tokenmax, ..."

Post by @cocoweixu on X — Wei Xu, NLP researcher, sharing a research update or announcement. Worth a second look for the underlying paper or result referenced.

Everything we launched during Agents Week — Our latest Agents Week has come to a close. Here’s a recap of all the announcements we made, from Wallets to Radar.

yc-software/qm: Multiplayer agent harness for work. — Multiplayer agent harness for work. Contribute to yc-software/qm development by creating an account on GitHub.

Agentic AI Summit 2026 - Center for Responsible, Decentralized Intelligence at Berkeley — UC Berkeley's RDI-hosted summit drew 5,000+ attendees and ~200 speakers across the AI agent stack. Speakers included Andrej Karpathy, Andrew Ng, and Berkeley researchers Sergey Levine, Ion Stoica, and Dawn Song.

Harvey introduces Tenet — Harvey's first model post-trained specifically for legal work, built on a Kimi K3 base. Post-training lifted its all-pass rate 82% on LAB and 22% on LAB Contracts versus the base model.

Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090 — Abstract page for arXiv paper 2608.27370: Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090

Redesigning the Inference Chip: From Nvidia GPU's Flaws to OpenAI Jalapeño — Deep dive on OpenAI's Jalapeño, a custom inference ASIC co-developed with Broadcom targeting ~50% lower cost per token than Nvidia GPUs. Explains how cutting L2 cache and wrapping HBM tightly around the compute die addresses inference's memory-bandwidth bottleneck.

AI Fluency for nonprofits — This course empowers nonprofit professionals to develop AI fluency in order to increase organizational impact and efficiency while staying true to their mission and values.

The most cited paper of the century is a brilliant hack — Welch Labs explores the technical breakthrough of residual networks and how skip connections solved the shattered gradient problem in deep learning. Deep Residual Learning for Image Recognition. This architectural innovation transformed neural networks, enabling unprecedented depth and shifting the fundamental understanding of how these models store and refine information during the learning process. Resource from the paper [A _Nature_ analysis evaluating the top-ranked papers published between 2000 and 2025](https://www.nature.com/articles/d41586-025-01125-9)

The Emergent Symbolic Structure of Artificial Neural Networks — McCoy, Soulos, Linzen, and Smolensky's paper (8 years in the making) showing neural networks implicitly learn structured, symbol-like representations resembling Tensor Product Representations. They replace a network's entire representation process with a closed-form symbolic equation and behavior stays largely unchanged.

Meta harness makes 10 times better Kimi K3 chip — A technical deep-dive rebuilding Moonshot AI's Kimi K3 inference chip architecture to hit ~10x higher clock frequency (100 MHz to 1047 MHz) using an open-source flow on Nangate45. The redesigned chip targets a ~1M-parameter "nano model," not the full 2.8T-parameter K3.

texts-to-transformer: Make Your Own Transformer — Must try. Train a small LLM from scratch on your own iMessage data, then run the trained model locally with MLX. Good candidate for writing a detailed junior-dev-style implementation plan (files to touch, what needs fixing) before diving in.

NVIDIA AI tweet — NVIDIA AI announcement/update shared on X. Check for latest NVIDIA AI product or research news.

Claude for Teachers - Anthropic — Anthropic's announcement of Claude for Teachers. New offering aimed at supporting educators with AI tools.

Legibility for Agents — Blog post on "legibility" as a design property for AI agents. Discusses making agent behavior and reasoning understandable to humans.

Qwen3-TTS-12Hz-1.7B-CustomVoice — Qwen's 1.7B parameter text-to-speech model with custom voice support at 12Hz. Open model on Hugging Face for TTS experimentation.

Artificial Analysis tweet — Update from Artificial Analysis, which tracks and benchmarks AI model performance. Likely new model comparison data.

openworker — Andrew Ng's open-source project for autonomous AI worker/agent workflows. Worth exploring for agent orchestration patterns.

The 2026 AI Engineering Report — Amplify Partners' survey of how AI engineering teams are actually building, evaluating, and shipping models this year.

GEPA for Structured Information Extraction for Enterprise Tasks — A DSPy tutorial showing GEPA-based prompt optimization applied to a real enterprise extraction task.

gbrain — Garry Tan's opinionated agent "brain" configuration for OpenClaw/Hermes-style autonomous agents.

The Agent-Native Document Editor — A document editor built from the ground up for AI agents to read, edit, and collaborate on, not just humans.

CoreWeave ARIA — Weights & Biases' agent framework built in partnership with CoreWeave for building and evaluating production agents.

MLE-bench — OpenAI's benchmark for evaluating how well AI agents perform real machine learning engineering tasks end-to-end.

GEPA: Optimize Anything with LLMs — An open framework that uses reflective prompt evolution to optimize arbitrary systems, not just prompts.

Ralph — An autonomous AI agent loop that keeps running against a PRD until every item on it is complete.

Healthcare x Real Estate: What’s happening in between them? — Can two clunky industries find something interesting in the middle?

Limbic Becomes First AI-Led Mental Healthcare Company Selected for FDA TEMPO — Limbic's clinician-supervised AI voice agent, Unpacked, joins the FDA's TEMPO pilot to deliver CBT to Medicare patients with depression or anxiety. First AI-led mental healthcare provider in the program.

FDA GenAI Framework — FDA guidance document on the use of generative AI in regulated healthcare and medical contexts. Relevant reference for AI safety and compliance work in health tech.

AI Reveals Hidden Patterns Inside Breast Cancer — Southampton researchers built CenSegNet, an open-source AI tool analyzing 330,000+ centrosomes across 911 breast cancer tumor specimens. Found two previously conflated forms of centrosome abnormality with implications for predicting tumor progression.

Rare Disease, Real Kid: The MVA Hackathon 2026 — Consider doing with Pallav. Hugging Face and Sage Bio hackathon offering $50K in prizes from Anthropic and AWS to analyze a real child's genome for a rare disease diagnosis; submissions close October 24.

Post by @biogerontology on X — Tempus trained a foundation model on 1.67M real cancer patients, fusing clinical trajectories with DNA, RNA, and H&E pathology into one pati [Paper](arxiv.org/abs/2608.24688) #aarohi

3rd Annual Statewide CHA & SMP Conference Week — 3rd Annual Statewide CHA & SMP Conference Week! Join our California Health Advocates Conference, Day at the Capitol, and California SMP Conference in Sacramento, CA January 11-15, 2027! Come together to learn, collaborate, advocate, and lead for a stronger Medicare system and a more secure future for all.

Arintra banks $25M series B funding round for AI-driven revenue assurance platform — Cofounder and CEO Nitesh Shroff told Fierce Arintra plans to use the funds to broaden its platform, expand to more enterprise health systems and more.

The data substrate for biological intelligence — AI-ready clinical data infrastructure for pharma, biotech, frontier AI labs, RWE teams, CROs, and health systems. 12.3M patients, five linked modalities, OMOP CDM normalized.

Aesto healthcare data breach impacts 9.5 million people — Healthcare firm Aesto has disclosed that a data breach announced earlier this year affected 9,540,683 individuals. #aws

A Reading Guide to Chinese Biotech — A curated primer for understanding China's fast-growing biotech sector and its key players.

preview.is RAG Search API — ** RAG/search API product concept for quickly adding retrieval backed search to applications. Interesting product idea for hosted RAG infrastructure and developer friendly search APIs.

Year of Systems — ** Ben Meer’s course/book companion built around installing one practical life system each week for a year. Useful for personal operations, habit infrastructure, and productivity systems.

“It’s Hard to Eval” Is a Product Smell — ** Hamel Husain argues that hard to evaluate AI products are often hard for users to trust or verify. Useful framing for AI product design, evals, and quality checks.