> ## Content Index > Fetch the complete content index at: https://secblog.reedcloudsec.com/llms.txt > Use this file to discover other available public pages before exploring further. # Vendor Watch — September 2026 - URL: https://secblog.reedcloudsec.com/vendor-watch/ - Published: 2026-09-16T02:42:00.000Z - Updated: 2026-09-16T02:42:00.000Z - Description: Two vendor announcements worth a skeptical read — one still just a roadmap item, one invite-only — plus real research: an AI-security technique crossing into phishing, and hard numbers on which edge vendors keep getting breached. - Author: Marselis A. Reed - Tags: Vendor Watch *Product launches, vendor research, and notable moves — read with the skepticism the hype usually deserves. secblog.reedcloudsec.com* Four items this round: two product announcements worth a critical read, and two genuine vendor research findings worth taking at face value. --- [**CrowdStrike Announces Agentic Identity Provider**](https://www.crowdstrike.com/en-us/blog/crowdstrike-announces-agentic-identity-provider/?ref=secblog.reedcloudsec.com) The problem CrowdStrike is naming here is real: AI agents today mostly run as service accounts or shared API keys, which makes it hard to tell which agent did what, or revoke one without breaking three others. The pitch is genuine per-agent identity — Falcon Guardian discovers agents across the environment, the Agentic Identity Provider registers each one with a cryptographically verifiable identity, then brokers short-lived, tightly scoped access per task instead of handing out standing credentials. Worth reading the actual language carefully, though: the announcement carries an explicit "unreleased services or features... subject to change" disclaimer, no GA date is given anywhere, and the technical detail stops well short of an architecture diagram or API spec. This is a problem worth solving, and CrowdStrike naming it this directly is a useful signal about where the industry's attention is heading — but right now it's a roadmap item, not something you can deploy. Worth tracking, not worth budgeting for yet. *via CrowdStrike Blog · Sep 2, 2026* [**Cloudflare: Context-Aware Vulnerability Discovery with Managed Defense and OpenAI Daybreak**](https://blog.cloudflare.com/vulnerability-discovery-remediation/?ref=secblog.reedcloudsec.com) The problem being solved is a real, common one: vulnerability scanners generate volume, not priority — a scanner will flag a finding without telling you whether that code path is even deployed or getting real traffic. Cloudflare's answer combines its own WAF and traffic telemetry (which routes are active, how much traffic they carry, and what security events surround them) with OpenAI's Daybreak model doing the actual reconnaissance and hunting against codebases you've explicitly authorized, then proposing either a code patch or a scoped WAF rule. One detail worth taking at face value, because it's a real architectural constraint rather than a marketing line: no model inference runs at Cloudflare's edge, and the model can't apply anything it proposes — a human reviews and deploys every patch and rule. That's the right design for this exact use case. The catch is availability: this is invitation-only early access through Cloudflare's Managed Defense team, not something you can sign up for today. *via Cloudflare Blog · Sep 3, 2026* [**Microsoft: ASCII Smuggling Crosses Over from AI Prompt Injection to Phishing Evasion**](https://www.microsoft.com/en-us/security/blog/2026/09/03/ascii-smuggling-crosses-over-from-ai-prompt-injection-to-phishing-evasion/?ref=secblog.reedcloudsec.com) Genuine vendor threat-intel, not a product pitch, and the technique is worth understanding on its own: ASCII smuggling hides invisible Unicode Tag-block characters (U+E0000–U+E007F) inside otherwise normal-looking text — a word like "funding" still reads as "funding" to a human, but an inserted invisible character breaks it apart for literal keyword-matching filters and can confuse ML tokenizers. The technique originally surfaced in AI prompt-injection research, hiding instructions from humans while keeping them machine-readable; Microsoft's telemetry shows it's now crossed over into plain phishing evasion. The campaign they tracked ran three months on a strict weekday-only schedule starting February 9, 2026, and peaked at 2.37 million messages in a single day. It used roughly 150 disposable finance-themed domains and relayed through the legitimate email platform ActiveCampaign to borrow its sender reputation. The honest finding, though: over 99% of these messages were still caught — not by Unicode-specific detection, but by the same layered defenses (sender reputation, URL analysis, authentication) that would have caught a conventional campaign. The lesson isn't "this technique is dangerous" — it's "don't rely on a single content-matching signal" — and Microsoft's own fix doubles as prompt-injection hardening: normalize and strip Unicode tag-block characters before any matching happens, upstream of both spam filters and any AI system that might otherwise read the hidden text. *via Microsoft Security Blog · Sep 3, 2026* [**SentinelOne: Edge Infrastructure Under Siege — What Two Independent Datasets Reveal**](https://www.sentinelone.com/blog/what-two-independent-datasets-reveal-about-whos-exploiting-your-perimeter/?ref=secblog.reedcloudsec.com) Genuinely useful cross-referenced research: SentinelOne compared Tenable's exposure telemetry (what's actually deployed and unpatched across real environments) against its own DFIR casework from 66 CVEs seen in actual breach investigations. The CVE-level overlap between the two datasets was surprisingly low — just 21% — but the vendor-level convergence was dramatic: 79% overall, and 100% among edge-specific vendors, with both teams independently landing on the same seven names: Fortinet, Citrix, Ivanti, Palo Alto Networks, Cisco, Juniper, and VMware. The individual numbers are worth sitting with: 54% of F5 customer environments contain an actively exploited CVE right now, and Citrix vulnerabilities take a median of 461 days to remediate. The most counterintuitive finding: high-priority CVEs take measurably longer to patch than lower-priority ones — a 24-day gap that runs exactly backwards from what a risk-based patching program should produce. Twelve CVEs in the dataset show confirmed multi-nexus attribution, meaning both state-sponsored and financially motivated actors independently found and exploited the same flaws. If your environment runs edge or perimeter gear from any of those seven vendors, this is less "interesting research" and more "audit your patch cadence against these specific numbers." *via SentinelOne Blog · Aug 26, 2026* --- *That's this round of Vendor Watch. secblog.reedcloudsec.com*