🔐 Cybersecurity Briefing — [Date]

Sources: The Hacker News · BleepingComputer · Krebs on Security · Have I Been Pwned · CISA · NVD

🔴 CRITICAL — Active Exploitation / High Priority

(Active exploitation, 0-days, major breaches, active campaigns)

Item: [Title] Category: [Label] Source: [Source Name] Summary: [1-2 sentence summary] URL: [Direct link]

🟠 HIGH — Vulnerabilities & Supply Chain

(Patchable vulnerabilities, supply chain attacks, malware)

Item: [Title] Category: [Label] ...

🟡 MEDIUM — Other News

(Policy changes, defensive tools, law enforcement actions, industry news)

Sources With No New Items (Past 24h)

Priority Ranking

``

If all sources produce nothing new in the past 24 hours, output exactly [SILENT] (nothing else). This is the cron job's no-content signal.

• RSS feed is the most reliable access method: https://www.bleepingcomputer.com/feed/
• The browser tool often hits Cloudflare "Just a moment..." — when this happens, fall back to RSS immediately
• RSS gives a good 1-2 sentence summary
• Articles from 22-23 Jun show good daily volume (5-8 articles/day)
• Sponsored content appears in the feed (e.g., "A Glimpse into the Search Your Target Market") — note as sponsored

The Hacker News

• Browser required — static curl returns empty (JS-rendered via Blogger)
• Articles have URLs like /2026/06/.html — date is in the URL
• Browser_console JavaScript extraction is the best approach
• ~5-8 articles per day, some are sponsored/ad content
• Has a "Weekly Recap" post every Monday — digests last week's news
• Note: thehackernews.uk subdomain is used for sponsored content — check the article link's hostname. If it ends in .uk, the content is sponsored/advertorial, not editorial. This is more reliable than date-based detection alone.

Krebs on Security

• Low volume — often 4-7 days between articles
• Most recent articles section: https://krebsonsecurity.com/
• Uses WordPress — dates are in elements inside
  tags on the main page, NOT on the

  tag itself or in entry-date spans
• Date extraction pitfall: The main page regex may return '?' for dates because the theme nests elements inside articles rather than using entry-date spans. When this happens, fall back to fetching individual article pages via curl -sL "" and extracting values
• Expect most runs to find nothing new

• NVD API 2.0 requires API key for reliable access (https://nvd.nist.gov/developers/request-an-api-key)
• Small unauthenticated queries often work — a 20-result, 24h window query (pubStartDate=...&pubEndDate=...&resultsPerPage=20) frequently succeeds without a key. Larger or historical queries will time out.
• Without key: web_search fallback for "recent critical CVEs", or use the small 24h API query above
• With key: https://services.nvd.nist.gov/rest/json/cves/2.0/?pubStartDate=...&pubEndDate=...&cvssV3Severity=CRITICAL
• Timeout with unauthenticated requests is common for queries larger than ~20 results or wider than 24h

• execute_code is BLOCKED in cron mode. The tool is disabled for cron jobs because it runs arbitrary local Python without user approval. ALL data parsing must use terminal() with inline Python (python3 -c "..." or heredocs python3 << 'PYEOF'). Never use execute_code in a cron workflow.
• NVD JSON may arrive double-escaped inside the tool-output wrapper from terminal() depending on the terminal backend. Direct json.loads() on piped stdin may fail. Detection: after curl -s "NVD_URL" -o /tmp/nvd_output.txt, try python3 -c "import json; json.load(open('/tmp/nvd_output.txt'))". If it parses, the JSON is normal. If it fails, it's double-escaped — use the recovery strategy in references/nvd-json-parsing.md. Always save NVD output to a file first (-o flag) so you can retry parsing either way without re-fetching.
• Do not fabricate CVSS scores. If NVD API fails, say "API unreachable" — don't guess.
• Do not fabricate article dates. If you can't extract a date, note it as "date unknown" rather than assuming it's recent.
• Do not fabricate article summaries. If the description is truncated, say so. Don't invent details.
• Empty result is valid. Many days have no major security news — output [SILENT] genuinely.
• Check for duplicate articles across sources (e.g., WhatsApp phishing reported by both THN and BleepingComputer). Merge into one entry with both sources listed.
• Time zone matters. BleepingComputer uses EDT (-0400), HIBP uses Z (UTC). Normalize to UTC for consistent 24h filtering. THN gives date only (no time) — treat as within 24h if the date matches today or yesterday.
• Sponsored content appears in feeds. Flag it as sponsored so the user can distinguish editorial vs. paid content.
• Cloudflare bypass is not always possible. When both browser and curl fail (e.g., BleepingComputer some days), use RSS feed. When RSS also fails, note the source as unreachable.
• Single browser session constraint. Only one browser_navigate session is active at a time — navigating to a new URL closes the previous page. Plan your browser workflow: fetch all curl/RSS sources in a single parallel batch first (using terminal calls), then do one browser session for all JS-rendered sites in sequence. Forced sequence example: navigate THN → extract → navigate CISA → extract → navigate Krebs (if curl failed) → extract.
• The Record RSS uses "GMT" timezone suffix. Python's datetime.strptime with %Z does not reliably parse "GMT". Always do pubDate.replace('GMT', '+0000') before parsing. Without this, items in the window will be silently missed because date parsing fails.
• CISA advisories and Krebs often have nothing new — don't slow down the whole pipeline waiting for them. Fetch them in the parallel batch and expect "no new items."
• SANS ISC RSS returns HTML instead of XML. The feed at https://isc.sans.edu/rssfeed.xml frequently serves the main page HTML regardless of Accept headers. When curl -sL returns , fall back to web_extract(urls=["https://isc.sans.edu/diary/0"]) — /diary/0 redirects to the most recent handler diary. Extract the publish date from the "Published: YYYY-MM-DD" line in the returned markdown.
• SANS ISC RSS pubDate uses "GMT" suffix (e.g. Wed, 01 Jul 2026 05:10:20 GMT), the same format as The Record. Python's strptime with %z or %Z does not reliably parse "GMT". Always do pubDate.replace('GMT', '+0000') before calling strptime — without this the date parse silently fails and entries are dropped from results.
• SANS ISC RSS is dominated by Stormcast podcast entries (daily "ISC Stormcast For Monday, ..."). Filter these out — they're podcast summaries, not written threat intelligence. The diary entries you want have substantive titles.
• THN sponsored articles have no explicit date marker in the h2 element. A missing date usually means the article is sponsored/advertorial — flag as sponsored rather than date-unknown.
• CISA advisory type extraction fails with selectors. The selectors .c-article__type, .field--name-field-advisory-type, and art.querySelectorAll('span') ALL return nothing on CISA's listing page. CISA renders advisory type (ALERT / PUBLICATION / EXTERNAL | RESOURCE) as a plain text node child of each
  , NOT inside a ,

  , or any other element. Always use art.textContent regex matching (e.g., (art.textContent || '').match(/\b(ALERT|PUBLICATION|EXTERNAL\s[|]\sRESOURCE)\b/)), never querySelectorAll('span').

Related

• blogwatcher skill — for RSS-feed-based blog monitoring with blogwatcher-cli
• references/threat-intel-deep-dive.md — deep-dive methodology for researching a specific CVE, ransomware operation, or threat actor. Structured analytical reports with attack chains, technical mechanisms, email delivery workflow, and the JADEPUFFER/CVE-2025-3248 research as a worked example. Use this when the user asks about a specific vulnerability or threat by name, rather than the periodic news roundup.
• references/dead-feeds-workarounds.md (under blogwatcher) — detailed curl/python3/browser fallback techniques
• references/api-queries.md (under security-briefing) — ready-to-use curl commands and python3 parsers for all 6 sources
• references/nvd-json-parsing.md (under security-briefing) — workaround for double-escaped JSON in cron mode terminal() calls
• cron-job-maintenance skill — for diagnosing cron job delivery/format issues

The user has provided the following instruction alongside the skill invocation: [IMPORTANT: You are running as a scheduled cron job. DELIVERY: Your final response will be automatically delivered to the user — do NOT use send_message or try to deliver the output yourself. Just produce your report/output as your final response and the system handles the rest. SILENT: If there is genuinely nothing new to report, respond with exactly "[SILENT]" (nothing else) to suppress delivery. Never combine [SILENT] with content — either report your findings normally, or say [SILENT] and nothing more.]

You are a security briefing agent. Gather threat intelligence from security blogs, CISA, NVD, vendor advisories, and AI security research.

Focus areas:

• AI/agent security vulnerabilities (prompt injection, MCP flaws, tool poisoning)
• CVEs affecting software Mike uses (Linux, Docker, nginx, Node, Python)
• Supply chain attacks and credential leaks
• Authentication/identity threats

Include a weekly StakeBench attack pattern update: check for new research on prompt injection vectors — indirect injection in web content, stealthy parasitism (attacks that succeed without disrupting user's task), multimodal injection (images carrying injected prompts), and agent architecture quirks that amplify risk (model choice, browser agent implementation). If any new StakeBench research or similar prompt injection benchmarks were published this week, summarize the attack patterns and success rates.

Format: brief digest with actionable items. Prioritize what affects self-hosted AI agents and small-scale infrastructure. Save to ~/.hermes/data/security-briefing/YYYY-MM-DD.md and write a short summary.

Response

Briefing saved to ~/.hermes/data/security-briefing/2026-08-12.md (164 lines, 7513 bytes). Here's the summary:

🔐 Cybersecurity Briefing — August 12, 2026

🔴 Top 6 Critical Items

1. LLM Reasoning Trace Exposure — API flaw in OpenAI/Anthropic/Google. Encrypted reasoning blocks replayed across sessions, decoded by weaker models. 315K blocks decoded, 367 PII + 182 credentials recovered. Already fixed by vendors — but review your logging/caching pipelines.

2. Microsoft Defender ShieldBreak Zero-Day — PoC claiming SYSTEM-level bypass of Defender. Dropped same day as Patch Tuesday.

3. VMware vCenter CVE-2026-59310 (CVSS 9.8) — Active exploitation with reverse_ssh cron persistence. Patch now if running vCenter.

4. CISA KEV — 3 new CVEs (CVE-2026-20349, CVE-2026-68820, CVE-2026-72898). Check your stack.

5. Adobe ColdFusion — Three CVSS 10.0 flaws (command injection, eval injection, auth bypass). Patch if in use.

6. Microsoft Patch Tuesday — 400 flaws, 3 zero-days (1 actively exploited Windows driver).

🟠 Supply Chain & Vulnerabilities

• LiteLLM / Trivy supply chain attack — 2,500+ orgs potentially exposed. Credential-stealing PyPI packages from March.
• Cisco ASA/FTD DoS — actively exploited
• SAP NetWeaver CVSS 9.8 — unauthenticated RCE
• Sandworm fake job interviews — trojanized WireGuard VPN
• SharePoint RCE now exploited in ransomware
• DeadLock ransomware using Polygon smart contracts

📊 StakeBench Update

The "Who Pays the Price?" paper (arXiv 2606.13385) evaluated 264 adversarial prompt injection cases across 4 agent configurations. Key finding: no attack objective is reliably resisted. Stealthy parasitism (attack succeeds without disrupting user task) is the most dangerous pattern. Also new: reasoning trace injection variant where a model's own thinking blocks are weaponized against it.

Your infra takeaway: Review agent tool access boundaries. Implement verification gates for high-value actions. The parts of the briefing that affect self-hosted AI agents directly are items 1, 7, 16, and the StakeBench section.