Daily · 2026-07-22

Ranked by engagement score at capture. Expand any card for full text; metrics are a flash read of importance.

Posts

6.2K

Σ score

Security

348K

Σ views

Product security & cyber · 2026-07-22

How urgency works · feed →

Text-derived urgency for triage — not CVSS/EPSS/KEV. Re-checked on every rebuild.

Critical exploit / RCE language High leak · bypass · ransomware Elevated clear security signal Needs score CVE, score next Watch lower-impact signal

Urgency (unverified)	Signal	CVE	Types
Elevated checked 2026-07-27	I wrote about the completely wild incident where OpenAI were testing a new model and it broke out of its sandbox and broke INTO Hugging F… 2026-07-22 · @simonw · next: `process_and_product_review`	`—`	agent_or_ai_risk, cybersecurity
Elevated checked 2026-07-27	Tucked away in this article is an appeal to the AI skeptics to PLEASE stop writing off stories like this OpenAI accidental exploit of Hug… 2026-07-22 · @simonw · next: `monitor`	`—`	product_security
Elevated checked 2026-07-27	wp2shell（CVE-2026-63030）を「入口で止められない」理由と、ランタイムで捕まえる設計 https://qiita.com/keitah/items/437ab259b8b52e8cf090 2026-07-22 · @HissyNC · next: `enrich_and_match_sbom`	`CVE-2026-63030`	cve, cybersecurity

Trend signals · last ~2 weeks

Weighted engagement (♥ + 3×↻ + 2×🔖 + views/500) by theme — points you can track over time. Hover dots for day detail.

Product securityCybersecurityOpen sourceDefense / dronesAgent infrastructurePrivacy / trustPractical tools

post preview

Cybersecurity @simonw 2026-07-22

♥784 ↻96 💬53 🔖601 👁210.6K score2.7K

I wrote about the completely wild incident where OpenAI were testing a new model and it broke out of its sandbox and broke INTO Hugging Face to steal the answers to the benchmark https://simonwillison.net/2026/Jul/22/ope…

Lesson: Independent writeups turn lab incidents into shared defender knowledge — don't dismiss as marketing.

Open on X link

Full post text

Product security @simonw 2026-07-22

♥1.2K ↻102 💬95 🔖177 👁111.4K score2K

Tucked away in this article is an appeal to the AI skeptics to PLEASE stop writing off stories like this OpenAI accidental exploit of Hugging Face as a dishonest marketing trick Frontier models can find and exploit vuln…

Lesson: Treat cyber-capable model incidents as real capability signals, not PR theater.

Open on X

Full post text

Open source @tom_doerr 2026-07-22

♥206 ↻18 💬2 🔖261 👁11.7K score805

Midday is an all-in-one business assistant for freelancers that integrates time tracking, invoicing, and financial management. It features automatic invoice-to-transaction matching via Magic Inbox and a secure vault for …

Lesson: Open-source ops agents (invoice, vault, matching) are the freelancers' ERP.

Open on X link

Full post text

Safe adopt prompt · repo / library / tool

Copy into Claude / Grok / Codex / Cursor — investigates provenance & malware first, then plans LifeOS-compatible install only if safe.

SAFE ADOPT — investigate first, then wire into my local agent stack (LifeOS-compatible)

You are my technical agent. Do NOT clone, download, pip/npm install, grant permissions, or write config until Phase A–D pass. If anything fails, STOP and report — never "install anyway."

══════════════════════════════════════
ARTIFACT (from Laitest capture)
══════════════════════════════════════
- Kind (guess): repo / library / tool
- Category: Open source (open_source)
- Source post: https://x.com/i/status/2080031409272410207
- Author on X: @tom_doerr (Tom Dörr) · date 2026-07-22
- Primary URL: https://github.com/midday-ai/midday
- All URLs:
  1. https://github.com/midday-ai/midday
- Capture lesson: Open-source ops agents (invoice, vault, matching) are the freelancers' ERP.
- Post excerpt:
"""
Midday is an all-in-one business assistant for freelancers that integrates time tracking, invoicing, and financial management. It features automatic invoice-to-transaction matching via Magic Inbox and a secure vault for storing contracts.

https://github.com/midday-ai/midday
"""

══════════════════════════════════════
GOAL
══════════════════════════════════════
Decide what this is (skill · tool · agent · MCP server · orchestrator · model · library · other), whether it is safe and worth adopting, and if yes: the minimal correct way to put it into MY stack so future sessions know how to use it.

My stack (LifeOS-compatible local agent OS):
- Skills live under ~/.claude/skills/ (use CreateSkill skill before creating/modifying ANY skill — never hand-roll SKILL.md)
- MCP servers / peer tools: configure only after provenance + security pass
- Projects/tools may land under ~/Projects/ or a clearly named path — never overwrite without asking
- Dual memory: curated notes vs verbatim recall — if worth remembering, say where it should be filed
- Prefer reversible, audited installs (git clone to a sandbox dir first; no curl|bash; no sudo unless I approve)

══════════════════════════════════════
PHASE A — IDENTIFY (no downloads yet)
══════════════════════════════════════
1. Name the artifact, license (if known), language/runtime, and intended use in one paragraph.
2. Classify exactly one primary type:
   [ ] LifeOS/Claude skill pack   [ ] MCP server   [ ] Coding agent/harness/CLI
   [ ] Library/SDK                [ ] Model/weights [ ] Orchestrator/workflow
   [ ] Security tool              [ ] Other: ___
3. List dependencies (Node/Python/Rust/Docker/GPU/API keys) and network blast radius (what it phones home to).
4. If multiple URLs, rank which is canonical and which are mirrors/docs/marketing.

══════════════════════════════════════
PHASE B — PROVENANCE (is this the real thing?)
══════════════════════════════════════
1. Resolve the canonical upstream (org/user + repo or package). Prefer official org over random forks.
2. Typosquat / clone checks:
   - Similar names with high stars that are NOT this one
   - Fork that rebrands itself as original
   - Recently created account + sudden stars (bought engagement)
   - README that copies a famous project but different remote
3. Author/reputation: real maintainers? Linked site? History of other repos?
4. Activity: last commit/release, open issues, PR response, abandoned?
5. Stars/forks/watchers in context (not gospel — only a signal).
6. Output: PROVENANCE = PASS | WEAK | FAIL with evidence links.

══════════════════════════════════════
PHASE C — SECURITY (malware / supply chain)
══════════════════════════════════════
Do this BEFORE any install that runs code:
1. Inspect README, install docs, GitHub Actions, package.json/pyproject/Cargo.toml for:
   - postinstall / prepare scripts that fetch remote code
   - obfuscated JS/Python, base64 blobs, eval/exec of network content
   - requests for secrets, SSH keys, wallet files, browser cookies
   - unexpected network endpoints, crypto miners, reverse shells
2. Check advisories: OSV / GitHub Dependabot / known CVE mentions for this package name.
3. Permissions: does it need root, full disk, browser automation, shell, or cloud credentials?
4. Sandbox plan: clone to a temporary directory (e.g. /tmp/laitest-vet-{slug} or ~/Projects/_quarantine/{slug}) with no global npm/pip install until review.
5. For models/weights: source integrity (checksums, HF org trust), license, data exfil risk if run with tools.
6. Output: SECURITY = PASS | CONDITIONAL | FAIL + specific red flags.
   CONDITIONAL requires listing exact mitigations (offline, container, no secrets, etc.).

══════════════════════════════════════
PHASE D — MAINTENANCE & FIT
══════════════════════════════════════
1. Is it maintained enough for production use, or research-only?
2. Does it duplicate something I already have? (search my skills/tools first if accessible)
3. Value vs complexity: is the lesson above real for me, or hype?
4. Output: FIT = ADOPT | WATCH | SKIP with one-sentence why.

══════════════════════════════════════
PHASE E — ONLY IF A–D allow it: SAFE ADOPT PLAN
══════════════════════════════════════
Produce a step-by-step plan I can approve:

1. Exact clone/download command into a quarantine path (no global side effects).
2. What to read first (entry files, SECURITY.md, LICENSE).
3. Minimal smoke test that does NOT need production secrets.
4. How to wire it into LifeOS / my agent of choice, choosing the right shape:
   - Skill → invoke CreateSkill; scaffold; do not hand-write SKILL.md
   - MCP → config snippet + scopes + why those permissions
   - CLI/agent → install location + PATH note + example invocation
   - Library → when to depend vs vendor
   - Model → where weights live, how to call, hardware needs
   - Orchestrator → how it coexists with existing harnesses (do not replace blindly)
5. Rollback: how to fully remove files, configs, services.
6. Memory note (2–4 bullets) for future sessions: what it is, when to use it, gotchas.

══════════════════════════════════════
OUTPUT FORMAT (strict)
══════════════════════════════════════
## Verdict
PROVENANCE / SECURITY / FIT — each PASS|WEAK|CONDITIONAL|FAIL|ADOPT|WATCH|SKIP

## What it is
…

## Red flags
… (or "none material")

## Safe next commands (only if not FAIL)
```bash
# quarantine only
```

## LifeOS / agent wiring (only if ADOPT)
…

## Memory note
…

## STOP reasons (if any)
…

Begin Phase A now. Do not skip to install.

Practical tools @tom_doerr 2026-07-22

♥105 ↻10 💬10 🔖113 👁8.1K score377

Access prediction market data and trades across Polymarket, Kalshi, and Limitless via the unified pmxt API. https://github.com/pmxt-dev/pmxt

Lesson: Unified prediction-market APIs let agents trade and research across venues.

Open on X link

Full post text

Access prediction market data and trades across Polymarket, Kalshi, and Limitless via the unified pmxt API. https://github.com/pmxt-dev/pmxt

Safe adopt prompt · repo / library / tool

Copy into Claude / Grok / Codex / Cursor — investigates provenance & malware first, then plans LifeOS-compatible install only if safe.

SAFE ADOPT — investigate first, then wire into my local agent stack (LifeOS-compatible)

You are my technical agent. Do NOT clone, download, pip/npm install, grant permissions, or write config until Phase A–D pass. If anything fails, STOP and report — never "install anyway."

══════════════════════════════════════
ARTIFACT (from Laitest capture)
══════════════════════════════════════
- Kind (guess): repo / library / tool
- Category: Practical tools (tools)
- Source post: https://x.com/i/status/2080055608737251659
- Author on X: @tom_doerr (Tom Dörr) · date 2026-07-22
- Primary URL: https://github.com/pmxt-dev/pmxt
- All URLs:
  1. https://github.com/pmxt-dev/pmxt
- Capture lesson: Unified prediction-market APIs let agents trade and research across venues.
- Post excerpt:
"""
Access prediction market data and trades across Polymarket, Kalshi, and Limitless via the unified pmxt API.

https://github.com/pmxt-dev/pmxt
"""

══════════════════════════════════════
GOAL
══════════════════════════════════════
Decide what this is (skill · tool · agent · MCP server · orchestrator · model · library · other), whether it is safe and worth adopting, and if yes: the minimal correct way to put it into MY stack so future sessions know how to use it.

My stack (LifeOS-compatible local agent OS):
- Skills live under ~/.claude/skills/ (use CreateSkill skill before creating/modifying ANY skill — never hand-roll SKILL.md)
- MCP servers / peer tools: configure only after provenance + security pass
- Projects/tools may land under ~/Projects/ or a clearly named path — never overwrite without asking
- Dual memory: curated notes vs verbatim recall — if worth remembering, say where it should be filed
- Prefer reversible, audited installs (git clone to a sandbox dir first; no curl|bash; no sudo unless I approve)

══════════════════════════════════════
PHASE A — IDENTIFY (no downloads yet)
══════════════════════════════════════
1. Name the artifact, license (if known), language/runtime, and intended use in one paragraph.
2. Classify exactly one primary type:
   [ ] LifeOS/Claude skill pack   [ ] MCP server   [ ] Coding agent/harness/CLI
   [ ] Library/SDK                [ ] Model/weights [ ] Orchestrator/workflow
   [ ] Security tool              [ ] Other: ___
3. List dependencies (Node/Python/Rust/Docker/GPU/API keys) and network blast radius (what it phones home to).
4. If multiple URLs, rank which is canonical and which are mirrors/docs/marketing.

══════════════════════════════════════
PHASE B — PROVENANCE (is this the real thing?)
══════════════════════════════════════
1. Resolve the canonical upstream (org/user + repo or package). Prefer official org over random forks.
2. Typosquat / clone checks:
   - Similar names with high stars that are NOT this one
   - Fork that rebrands itself as original
   - Recently created account + sudden stars (bought engagement)
   - README that copies a famous project but different remote
3. Author/reputation: real maintainers? Linked site? History of other repos?
4. Activity: last commit/release, open issues, PR response, abandoned?
5. Stars/forks/watchers in context (not gospel — only a signal).
6. Output: PROVENANCE = PASS | WEAK | FAIL with evidence links.

══════════════════════════════════════
PHASE C — SECURITY (malware / supply chain)
══════════════════════════════════════
Do this BEFORE any install that runs code:
1. Inspect README, install docs, GitHub Actions, package.json/pyproject/Cargo.toml for:
   - postinstall / prepare scripts that fetch remote code
   - obfuscated JS/Python, base64 blobs, eval/exec of network content
   - requests for secrets, SSH keys, wallet files, browser cookies
   - unexpected network endpoints, crypto miners, reverse shells
2. Check advisories: OSV / GitHub Dependabot / known CVE mentions for this package name.
3. Permissions: does it need root, full disk, browser automation, shell, or cloud credentials?
4. Sandbox plan: clone to a temporary directory (e.g. /tmp/laitest-vet-{slug} or ~/Projects/_quarantine/{slug}) with no global npm/pip install until review.
5. For models/weights: source integrity (checksums, HF org trust), license, data exfil risk if run with tools.
6. Output: SECURITY = PASS | CONDITIONAL | FAIL + specific red flags.
   CONDITIONAL requires listing exact mitigations (offline, container, no secrets, etc.).

══════════════════════════════════════
PHASE D — MAINTENANCE & FIT
══════════════════════════════════════
1. Is it maintained enough for production use, or research-only?
2. Does it duplicate something I already have? (search my skills/tools first if accessible)
3. Value vs complexity: is the lesson above real for me, or hype?
4. Output: FIT = ADOPT | WATCH | SKIP with one-sentence why.

══════════════════════════════════════
PHASE E — ONLY IF A–D allow it: SAFE ADOPT PLAN
══════════════════════════════════════
Produce a step-by-step plan I can approve:

1. Exact clone/download command into a quarantine path (no global side effects).
2. What to read first (entry files, SECURITY.md, LICENSE).
3. Minimal smoke test that does NOT need production secrets.
4. How to wire it into LifeOS / my agent of choice, choosing the right shape:
   - Skill → invoke CreateSkill; scaffold; do not hand-write SKILL.md
   - MCP → config snippet + scopes + why those permissions
   - CLI/agent → install location + PATH note + example invocation
   - Library → when to depend vs vendor
   - Model → where weights live, how to call, hardware needs
   - Orchestrator → how it coexists with existing harnesses (do not replace blindly)
5. Rollback: how to fully remove files, configs, services.
6. Memory note (2–4 bullets) for future sessions: what it is, when to use it, gotchas.

══════════════════════════════════════
OUTPUT FORMAT (strict)
══════════════════════════════════════
## Verdict
PROVENANCE / SECURITY / FIT — each PASS|WEAK|CONDITIONAL|FAIL|ADOPT|WATCH|SKIP

## What it is
…

## Red flags
… (or "none material")

## Safe next commands (only if not FAIL)
```bash
# quarantine only
```

## LifeOS / agent wiring (only if ADOPT)
…

## Memory note
…

## STOP reasons (if any)
…

Begin Phase A now. Do not skip to install.

post preview

Cybersecurity @HissyNC 2026-07-22

♥33 ↻7 💬1 🔖29 👁3.9K score120

wp2shell（CVE-2026-63030）を「入口で止められない」理由と、ランタイムで捕まえる設計 https://qiita.com/keitah/items/437ab259b8b52e8cf090

Lesson: Some WordPress RCEs bypass WAF at the edge — design for runtime detection, not perimeter-only.

Open on X link

Full post text

wp2shell（CVE-2026-63030）を「入口で止められない」理由と、ランタイムで捕まえる設計 https://qiita.com/keitah/items/437ab259b8b52e8cf090

Open source @itsharmanjot 2026-07-22

♥30 ↻13 💬2 🔖22 👁2.3K score118

Repos about to blow up: Bumblebee (MCP security workflows), Hyperframes (HeyGen OSS — agent writes HTML → MP4), OpenMythos, map-anything (single-pass 3D, Apache 2.0), Camofox Browser (production engine open-sourced).

Lesson: Agent-native media (HTML→video) and local MCP security scanners are rising together.

Open on X

Full post text

Lessons learned

Independent writeups turn lab incidents into shared defender knowledge — don't dismiss as marketing.
Treat cyber-capable model incidents as real capability signals, not PR theater.
Open-source ops agents (invoice, vault, matching) are the freelancers' ERP.
Unified prediction-market APIs let agents trade and research across venues.
Some WordPress RCEs bypass WAF at the edge — design for runtime detection, not perimeter-only.
Agent-native media (HTML→video) and local MCP security scanners are rising together.