Hey there! As mentioned, I spent last weekend running a bunch of evals, and tl;dr is I’ll absolutely be investing more in GLM 5.3 as an option for routing my work through until we can get access to models with classifier restrictions lifted.
Related: there’s been a bunch of news about OpenCode this week, including access to free models. I haven’t looked into this much yet, but it could be an option for getting up and running with building a hackbot at low/no cost. But they also just got hacked and leaked a bunch of user’s keys… sooo.
In personal news, I’m getting ready to head out on vacation for the next couple of weeks, so you may not see me publish (or maybe you will… from a beach. Depends on how I’m feeling lol).
Anyway, let’s dive in.
Nothing to promote. Just happy you’re here!
OpenCode's Big Week Ends in a Breach: Free Models, 2.0, Then Stolen Provider Keys [𝕏 Tweet]
by OpenCode
OpenCode spent the week giving away frontier models (DeepSeek V4.1 Flash, GLM 5.3 Flash, Nemotron 3.5 Super VL, Beam with 1M context) and shipping OpenCode 2.0, then disclosed a breach tied to a large-scale distillation attack. It hit 1.4% of OpenCode Console users with external providers linked: their stored OpenAI, Anthropic and Google keys were used to run up charges. Affected users were emailed, refunds are underway, and OpenCode is working with the labs to remediate. Rotate any provider key you ever saved in OpenCode.
PwnEscape: Linux Networking Bug Behind the Vercel Sandbox KVM Escape (CVE-2026-99999) [📓 Blog]
by Paulos Yibelo (@PaulosYibelo)
Paulos published the PwnEscape write-up detailing a Linux networking flaw that can lead to KVM guest-to-host escapes in sandboxed environments. The post includes the technical context and mitigation guidance relevant to teams running affected hypervisor configurations.
Have something you want to Spotlight? Tell me.
Intigriti on Earlier ID Verification for Public Programs [📓 Blog]
by Intigriti
Intigriti outlines why some platforms are moving ID verification earlier in the submission flow to reduce noise and improve triage signal. The post also calls out the trade-off: earlier verification can exclude researchers who rely on anonymity, making strong responsible-disclosure paths and privacy safeguards more important.
HackerOne Cuts Trial Reports to 3 for Programs You Don't Yet Qualify For [📓 Blog]
by HackerOne
HackerOne changed how trial reports work between Oct 12 and 19: every hacker now gets 3 trial reports for programs whose Signal requirement they don't meet, down from separate per-program and platform-wide allowances. Any triaged or resolved report (VDP or BBP) resets the limit to 3, a duplicate of an undisclosed triaged report returns one, and otherwise there's a 90-day pause after the third.
Chrome VRP Restructures Rewards (Higher Memory-Safety Payouts, Fuzzer Bonus) [📓 Blog]
by Google Bug Hunters · written by Alex Gough, Jon Serpas
Google Bug Hunters announced a Chrome VRP revamp aimed at rewarding deeper work, including significantly higher memory-safety base rates with process-specific caps and updated flat rates for several classes. The update also adds a $2,000 flat rate for S2 non-memory issues that lead to a fix, increases potential payouts for non-memory sandbox escapes, and revives a $2,000 Chrome Fuzzer Bonus; submissions queued after April 30, 2026 are evaluated under the new rules.
HackerOne Benchmarks 18 Frontier LLMs on 62 Complex Vulnerabilities [📓 Blog]
by HackerOne · research by Devansh (@0xAsm0d3us)
HackerOne published results from a benchmark comparing 18 frontier LLMs against 62 complex vulnerability cases, focusing on vulnerability understanding and analysis performance. The write-up shares comparative outcomes intended to help teams reason about model strengths and limitations in security triage and vuln research workflows.
Gray Swan Launches $300K “IPI Black” Prompt-Injection Challenge [📓 Blog]
by Gray Swan
Gray Swan opened a $300,000 prompt-injection challenge (IPI Black) focused on finding exploitable injection paths in LLM workflows. The program is backed by Anthropic, Meta, and Google DeepMind, signaling continued momentum toward structured, competitive testing for agent and prompt-injection failure modes.
Meta Bug Bounty Joins NahamCon 2026 Prompt2Pwn Live Hacking [𝕏 Tweet]
by Ben Sadeghipour (@NahamSec)
Ben announced Meta Bug Bounty’s participation in NahamCon 2026, including the Prompt2Pwn live hacking event. The post points to event details for the in-person session.
SPIRITCYBER26 Wraps Up With Singapore CTF Winners Announced [𝕏 Tweet]
by YesWeHack
YesWeHack announced the winners of SPIRITCYBER26, with top teams including nullIoT, One Prompt to Pwn Them All, Peenoise, Hamtarou, and Undefined Behavior. The post also highlights collaboration with CSA Singapore around security for connected technologies supporting the Smart Nation initiative.
Did I miss an important update? Tell me.
BB Arena: Black-Box Benchmark for AI Agents Finding Real Web/API Bugs [🛠️ Tool]
by Marius du Preez (@mdp_sec)
Marius released BB Arena, a black-box benchmark that evaluates autonomous AI agents against 100 private web/API labs, requiring discovery through proof-of-exploit to count as solved. The site publishes per-model solve rates, cost-per-solve, and first-attempt stats, with methodology designed to reduce contamination from training data.
Also: v2 / open-sourcing v1
Claude Kit: Claude Code Plugin for Triager-Grade Bug Bounty Reports [🛠️ Tool]
by YesWeHack
YesWeHack’s Claude Kit is a Claude Code plugin designed to produce triager-friendly reports strictly from user-provided facts, with guardrails intended to reduce hallucination and impact overclaims. It adds workflows like /ywh:write for report drafting, /ywh:triage for submission readiness (READY / NEEDS FIXES / DO NOT SUBMIT), and /ywh:gotchas for vulnerability-class-specific proof expectations and common false positives.
rust_ASC: Fast, Low-Latency Android Decompiler for Static Analysis [🛠️ Tool]
by MG1937
rust_ASC is a Rust reimplementation of ASC focused on low-latency, memory-efficient static analysis of APKs, with support for multi-DEX and modern DEX container formats. The repo includes search and decompile primitives plus benchmarks showing large performance gains over the Python reference and strong parallel scaling for full-APK export workflows.
Have a favorite tool? Tell me.
Adobe Experience Manager Cloud: Pre-Auth RCE via Vulnerability Chaining [📓 Blog]
by Searchlight Cyber (@SLCyberSec) · research by Dylan Pindur
Dylan details an unauthenticated RCE chain against Adobe Experience Manager (AEM) Cloud by combining multiple issues, including a DAM bulk import flaw, stored XSS via DownloadServlet, XXE via CreateCoralTreeServlet, and Groovy execution via /bin/asynccommand. The write-up also covers dispatcher bypass techniques, JCR query tactics for identifying accessible nodes, and notes AI-assisted auditing (Opus 4.6) as part of the workflow.
Visual Studio Copilot & SSMS RCE via Trailing-Dot Path Canonicalization (CVE-2026-103572) [📓 Blog]
by Johann Rehberger
Johann breaks down CVE-2026-103572, an RCE in Microsoft.VisualStudio.Copilot.Core.dll affecting Visual Studio and SQL Server Management Studio, driven by filename validation bypasses using trailing dots and Win32 canonicalization differences. The post shows how this can lead to overwriting sensitive configuration (including MCP server config and ripgrepPath), includes PoCs and timeline details, and notes Microsoft’s fix via additional GetFullPath normalization.
$90K Google DevSite RCE: Signed Flask Session Cookie Leads to Pickle Deserialization [📓 Blog]
by Simon Koeck
Simon reports an RCE in Google DevSite where a signed Flask session cookie contained a zlib-compressed pickled oauth2client OAuth2WebServerFlow object, including a client_secret. Because that secret was reused to derive the cookie-signing key, forged sessions enabled arbitrary pickle injection and server-side code execution; the write-up includes a PoC and remediation guidance around secret handling and avoiding pickle for client-controlled data.
Breaking into Google’s GFile for $100K [📓 Blog]
by Arvin Shivram
Arvin details a $100K Google bounty centered on exploiting GFile, including the investigation path and the technical lessons that made the bug report actionable. The post focuses on how the underlying behavior was proven and communicated in a way that held up through triage.
Symfony Cache-Control ReDoS (CVE-2026-108082): Seconds of CPU per Request [📓 Blog]
by Martino Spagnuolo (@r3verii)
Martino documents CVE-2026-108082, a regex-based ReDoS in symfony/http-foundation’s HeaderUtils::split() reachable via a malicious Cache-Control header. The post shows how small payloads can trigger catastrophic backtracking (multi-second CPU burn per request), discusses delivery constraints across common proxies/CDNs, and lists fixed versions (6.4.48 / 7.4.21 / 8.1.9) plus WAF-style mitigations.
Did I miss something? Tell me.
Codex vs Claude Code vs Gemini CLI: Where Each Agent Fits in a Bug Bounty Workflow [📓 Blog]
by YesWeHack
YesWeHack compares Claude Code (Opus 4.8), Codex (GPT-5.6 Sol), and Gemini CLI (Flash-3.5) on PortSwigger mystery labs, highlighting distinct strengths across recon, reasoning, and autonomous validation. The post includes concrete lab outcomes (DOM XSS, H2.TE request smuggling, JWT forgery, Java deserialization) and emphasizes human-verified, reproducible evidence—especially when agents cross from analysis into active testing.
Sn0rkY (Joffrey Czarny) on Bug Bounty: Deep Understanding Over Tool Reliance [📓 Blog]
by YesWeHack
In this YesWeHack interview, Sn0rkY (Joffrey Czarny) describes an approach centered on understanding systems deeply, building custom labs, and prioritizing curiosity over heavy automation. He frames AI as potentially useful but not essential to his workflow, and shares newcomer advice around sticking with targets long enough to find non-obvious logic bugs instead of program-hopping.
HackerNotes Ep. 196: Bug Bounty, Pentesting, and Local Models with rez0 and gr3pme [📓 Blog]
by Critical Thinking Podcast · written by Aituglo
This episode write-up covers rez0 and gr3pme’s discussion of how local models and agent orchestration are changing pentesting and bug bounty economics, including the “refusal tax” of gated model tiers and the operational benefits of cheap local tokens. It argues that specificity and scoped tasking are increasingly important for getting useful outputs, while forecasting fewer low/medium findings and more emphasis on deeper research.
Did I miss something? Tell me.
AI Hacking: Bug Bounty Edition (DEF CON 34 BBV Workshop) [🎥 Video]
by Bug Bounty Village · talk by Ben Sadeghipour (@NahamSec) and Clovis Mint
This DEF CON Bug Bounty Village workshop covers attack patterns against agentic AI systems, led by Ben and Clovis. It walks through challenge-style examples including system prompt leakage, tool-abuse paths that resemble IDOR-style outcomes, and indirect prompt injection via poisoned data sources, with reporting/validation guidance for making impact legible to triage.
Local AI: The Future of Pentesting & Bug Bounty? (Ep. 196) [🎥 Video]
by Critical Thinking Podcast
This episode features Rhynorater, rez0, and gr3pme discussing how local models and agent workflows are changing pentesting throughput and bug bounty competitiveness. It covers practical architecture choices for running local setups and where the panel thinks high-leverage opportunities may remain as automation improves.
Building a Hackbot for Under $250 [🎥 Video]
by Ben Sadeghipour (@NahamSec)
Ben shows a budget “hackbot” build-out on a low-cost VPS for continuous recon and initial triage. The video covers provisioning and wiring a recon pipeline (discovery, scanning, probing, templated checks), notification routing, and operational considerations, with a concrete cost breakdown.
How Hackers Build an AI Advantage (with Douglas Day) [🎥 Video]
by HackerOne
HackerOne interviews Douglas Day on how offensive researchers are integrating AI into recon, triage, and chaining workflows. The discussion also covers practical pitfalls like prompt injection and why human validation remains necessary for business-logic-heavy findings.
AI Hackbot Recon Workflow: Subdomains to Findings [🎥 Video]
by zack0x01
zack0x01 walks through an automated recon workflow for an AI-assisted “hackbot,” from subdomain collection through turning recon output into candidate findings. The video focuses on end-to-end pipeline design and how results are surfaced for manual follow-up and validation.
Did I miss something? Tell me.
New XSS Vectors Using <audio>/<video> and onwebkitneedkey [𝕏 Tweet]
by Gareth Heyes (@garethheyes)
Gareth shared two XSS vectors using <audio>/<video> with the onwebkitneedkey handler, added to PortSwigger’s XSS cheat sheet. The examples include a PoC demonstrating Safari-triggered execution.
Using Historical Bug Reports as AI “Training Data” for Non-Duplicate Hunting [𝕏 Tweet]
by Critical Thinking Podcast
Critical Thinking Podcast suggests seeding an AI workflow with prior bug reports and internal techniques so it searches for similar behaviors in new targets, instead of running generic scans that tend to duplicate. The idea is framed as a way to bias tooling toward program-specific, less-commoditized findings.
Akamai WAF XSS Bypass Payload: SVG onload [𝕏 Tweet]
by 𝐤𝐣𝐮𝐥𝐢𝐮𝐬 (@ethical_h4ck3r_)
A compact SVG onload payload claimed to bypass Akamai WAF and trigger XSS. Useful as a quick regression test when probing WAF handling of SVG.
AI Agent Pulls 10K+ PII Records from Old JavaScript Artifacts [𝕏 Tweet]
by Ben Sadeghipour (@NahamSec)
Ben describes using an AI agent to scan older CaidoIO JavaScript files and extract over 10,000 PII records (including names, DOBs, and emails). The post outlines a simple three-step workflow for mining stale frontend artifacts for sensitive data exposure.
$100 Priority Triage Fee to Reduce Backlog and Speed Payouts [𝕏 Tweet]
by Sam Curry (@samwcyo)
Sam proposes a $100 per-report fee for priority triage, automated duplicate checks, and payouts under 48 hours, citing positive experience with similar dynamics on Immunefi. The suggestion is positioned as a way to improve throughput and reduce the operational burden of low-signal submissions.
Did I miss something? Tell me.
Because Disclosure Matters: This newsletter was produced with the assistance of AI. While I strive for accuracy and quality, not all content has been independently vetted or fact-checked. Please allow for a reasonable margin of error. The views expressed are my own and do not reflect those of my employer.