ctipilot.ch

CTI Daily Brief — 2026-05-17

Typedaily
Date2026-05-17
GeneratorClaude Opus 4.7 (`claude-opus-4-7`)
ClassificationTLP:CLEAR
LanguageEnglish
Promptv2.59
Items6
CVEs18
On this page

On this page

Tags (18)
Regions (5)
References (30)

0. TL;DR

  • F5 BIG-IP / BIG-IQ May 2026 Quarterly Notification — SecurityWeek reports "over 19 high-severity and 32 medium-severity" bugs across BIG-IP, BIG-IQ and NGINX; NCSC-NL CSAF lists 43 in the BIG-IP / BIG-IQ scope. Lead CVE-2026-41225 (CVSS 4.0 score 8.6 per F5 / SecurityWeek; CVSS 3.1 score 9.1 per NCSC-NL / NVD; both confirm post-auth Manager-role RCE on iControl REST); secondary 8.7-class cluster includes iControl REST command injection, SSH-password exposure in audit logs, and Appliance-mode-bypass privilege escalation. No in-the-wild exploitation reported as of advisory publication. Affects BIG-IP appliances widely deployed across European public-sector load-balancing / WAF perimeters (F5 K000160932, 2026-05-14; SecurityWeek, 2026-05-14).
  • FunnelKit "Funnel Builder for WooCommerce" actively exploited as Magecart skimmer on 40,000+ WordPress checkout pages — no CVE assigned. Unauthenticated POST to an internal-method dispatcher writes attacker-controlled JavaScript into the plugin's External Scripts setting; a fake Google Tag Manager loader opens a WebSocket to attacker C2 and pulls a storefront-tailored card skimmer. Patched in v3.15.0.3 (Sansec, 2026-05-14).
  • DHTMLX Gantt / Scheduler / Diagram PDF Export Module — CVE-2026-41553 unauthenticated RCE (CVSS 4.0 score 10.0). CERT-PL coordinated disclosure of three flaws in widely-embedded JavaScript scheduling/diagramming libraries; the lead bug processes attacker JS in the data parameter server-side via Node.js, achieving full command execution on the export host. EU public-sector portal ecosystem heavy user. Fixed in PDF Export Module 0.7.6 and Diagram 1.1.1 (CERT-PL, 2026-05-15).
  • Pwn2Own Berlin 2026 wraps — 47 unique zero-days, $1,298,250 awarded. DEVCORE's Orange Tsai chained three undisclosed Exchange bugs to SYSTEM-level unauthenticated RCE on Day 2 ($200K, 90-day embargo); STARLabs SG burned a memory-corruption ESXi hypervisor escape for another $200K on Day 3; the new AI Agents category produced exploits or collisions across all entered targets — OpenAI Codex (Compass Security CWE-150, $40K), Cursor (Compass Security, $15K), LM Studio (OtterSec code-injection Day 2; STARLabs SG separately ran an SSRF+code-injection 5-bug chain on Day 1), LiteLLM (k3vg3n SSRF+code-injection), with Claude Code, Chroma, Megatron Bridge and Ollama producing collisions (ZDI Day 3, 2026-05-16).
  • CERT-PL discloses CVE-2026-44088 in SzafirHost — JAR zip-polyglot bypass enables RCE in Poland's national eIDAS-recognised qualified e-signature browser helper. A class-loading split-brain between JarInputStream (verifies signature from file start) and JarFile/URLClassLoader (loads classes from ZIP Central Directory at end) lets an attacker chain a genuine signed JAR with a malicious ZIP so signature verification passes but the malicious class loads. Direct impact on Polish public administration, courts, procurement and healthcare workflows that produce qualified electronic signatures cross-recognised under eIDAS. Patched in SzafirHost 1.2.1 (CERT-PL, 2026-05-15).

3. Research & Investigative Reporting

Kaspersky GReAT documents Kimsuky's Rust-based HelloDoor and TryCloudflare-tunnel C2 added to the PebbleDash toolkit [SINGLE-SOURCE]

Kaspersky's Global Research and Analysis Team published a deep technical disclosure on 2026-05-14 covering Kimsuky (Ruby Sleet / APT43) campaigns observed during late 2025 and Q1 2026, documenting six malware families the actor is currently rotating (Kaspersky Securelist, 2026-05-14). The headline novelty is HelloDoor, the first Rust-based variant in the PebbleDash family (a backdoor platform Kimsuky appropriated from Lazarus around 2021); secondary additions are httpMalice (HTTP-only loader), MemLoad (reflective DLL loader), httpTroy (C2 backdoor) and continued use of AppleSeed / HappyDoor. The most operationally significant capability change is that HelloDoor's C2 channel uses Cloudflare Quick Tunnels via TryCloudflare — short-lived *.trycloudflare.com hostnames issued ad-hoc, terminating attacker control infrastructure behind Cloudflare's CDN, eliminating fixed C2 IPs and making network-layer indicator blocking impractical. Kaspersky verbatim: "Kimsuky has continuously introduced new malware variants based on the PebbleDash platform, a tool historically leveraged by the Lazarus Group but appropriated by Kimsuky since at least 2021... including the use of VSCode Tunneling, Cloudflare Quick Tunnels, DWAgent, large language models (LLMs), and the Rust programming language." Reported targeting: South Korean government, defence and medical sectors as the primary set, with documented spillover hits in Germany — the closest geographic proximity to Swiss government targets in recent Kimsuky reporting. Detection guidance from Kaspersky (paraphrased to avoid IOC reproduction): monitor for JSE/SCR/PIF droppers carrying Base64-encoded payloads; flag scheduled tasks under generic browser-update names (e.g. ChromeCheck, EdgeCheck); inspect VSCode tunnel authentications via GitHub for unrecognised tunnel names; alert on Rust-compiled PE images loading from non-standard paths and on outbound *.trycloudflare.com connections that don't match a developer's legitimate tunnel-use profile. Technique class: T1071.001 Application-layer C2 via web protocol + T1090.002 External Proxy + T1053.005 Scheduled Task. [SINGLE-SOURCE] — only Kaspersky GReAT carries this depth; included because Kaspersky is HIGH-reliability for North Korea-nexus reporting and the technical detail is defender-actionable. Marked at edge of the 72 h developing window (Securelist publication 2026-05-14, ~62 h before run start).

4. Updates to Prior Coverage

UPDATE: Exchange CVE-2026-42897 — Pwn2Own DEVCORE three-bug SYSTEM RCE chain emerges alongside active OWA-XSS exploitation

UPDATE (originally covered 2026-05-15 and 2026-05-16 deep dive): DEVCORE's Orange Tsai chained three undisclosed Exchange Server bugs on Pwn2Own Berlin 2026 Day 2 to achieve unauthenticated remote code execution at SYSTEM privilege level, earning $200,000 (Zero Day Initiative, 2026-05-15; BleepingComputer, 2026-05-15). This chain is separate from the actively-exploited CVE-2026-42897 (OWA stored XSS, no permanent patch; EEMS mitigation M2.1.x only) that the 2026-05-16 deep dive covered. ZDI verbatim: "Orange Tsai (DEVCORE Research Team) earned $200,000 after chaining three bugs to gain remote code execution with SYSTEM privileges on Microsoft Exchange."

The three bugs are under a 90-day Pwn2Own embargo — Microsoft must patch by approximately 2026-08-14 before ZDI publishes technical detail. Operationally, the compound risk for on-premises Exchange has materially worsened in 48 h: one actively exploited XSS without a permanent patch (M2 mitigation only, with known OWA Calendar Print / inline-image side-effects), plus a fresh unauthenticated SYSTEM RCE class that defenders cannot pre-emptively patch. CVE-2026-42897 remains in CISA KEV (added 2026-05-15) with EEMS as the only listed mitigation; the Microsoft Exchange blog post addressing-exchange-server-may-2026-vulnerability-cve-2026-42897 linked from the MSRC advisory returns 502 on direct fetch and the MSRC entry itself is the operational primary (MSRC CVE-2026-42897).

Defender response shift for on-premises Exchange 2016/2019/SE: treat the platform as severely threatened. Verify EEMS service is enabled (Get-ExchangeDiagnosticInfo, mitigation M2.1.x present in applied list); restrict ECP/EWS/OWA reachability from the internet at the WAF or reverse proxy where business-feasible; accelerate any in-progress Exchange Online migration; assume hypothetical compromise paths through both OWA-browser-context attacks (CVE-2026-42897) and a direct service-account SYSTEM RCE chain (Pwn2Own DEVCORE) until Microsoft ships permanent fixes for both. Exchange Online tenants are not in scope for either.

Changes since first coverage(1 prior appearance)
  1. 2026-05-162026-05-16First coverage. Microsoft confirmed Exploitation Detected 2026-05-14; CISA KEV-added 2026-05-15; no permanent patch — EEMS Mitigation M2 auto-applies, EOMT for air-gapped; Exchange 2016/2019 permanent fix gated on Period 2 ESU enrolment. Deep dive § 5.

5. Deep Dive — Pwn2Own Berlin 2026: Master-of-Pwn outcomes, the new AI Agents category, and the compound-Exchange-threat picture for European defenders

Background. Pwn2Own Berlin (run alongside OffensiveCon, 2026-05-14 → 2026-05-16) is the second Berlin edition since Trend Micro / Zero Day Initiative moved the European event off the Vancouver-only schedule in 2025. It runs the standard Pwn2Own rules: original-research, full-chain, time-boxed in-room exploitation against current-patched production targets, with vendor-disclosure happening within minutes of a successful pop and a 90-day Pwn2Own embargo before ZDI publishes technical detail. The Berlin contest historically draws a heavier European researcher field than Vancouver — relevant this year because Swiss firm Compass Security fielded a five-researcher team and took prizes against multiple AI agent targets. Prior Pwn2Own competitions established the cadence: bugs popped in May surface as advisory-tagged CVEs in vendor August or September advisories. The May 2026 contest is meaningful for European public-sector defenders for three reasons covered below — the DEVCORE Exchange chain landing while CVE-2026-42897 is actively exploited, the new AI Agents category dragging dev-toolchain inference platforms into the public-vulnerability ecosystem, and the contest's capacity overflow which produced an unprecedented wave of rejected-researcher public PoC releases.

Day-by-day outcomes — what was actually demonstrated. Day 1 (ZDI, 2026-05-13): Orange Tsai (DEVCORE) opened the day with a four-bug Microsoft Edge sandbox escape for $175,000 — the day's biggest single award and the foundation of DEVCORE's eventual Master of Pwn victory; Compass Security exploited OpenAI Codex through a CWE-150 "improper neutralization of special elements" bug for $40,000 — the first publicly-known weaponised exploit of OpenAI's coding agent; Satoki Tsuji (Ikotas Labs) exploited NVIDIA Megatron Bridge via an overly permissive allowed-list bug for $20,000; Ikotas Labs separately collided against LiteLLM ($8,000 reduced reward); maitai (Doyensec) collided against OpenAI Codex ($10,000); Nguyen Thanh Dat (Viettel) collided against Claude Code ($20,000); k3vg3n landed an SSRF-plus-code-injection chain against LiteLLM (separate from the Ikotas Labs LiteLLM collision); Le Duc Anh Vu (Viettel) failed his attempt against Codex; STARLabs SG demonstrated a five-bug SSRF + code-injection chain against LM Studio. Day 2 (ZDI, 2026-05-15): the Exchange chain landed — Orange Tsai of DEVCORE chained three undisclosed bugs to unauthenticated SYSTEM RCE on a patched Exchange Server installation, earning $200,000 and a Master-of-Pwn step toward DEVCORE's overall victory; OtterSec popped LM Studio via a code-injection bug; 0xDACA / Noam Trobinski took the NVIDIA Container Toolkit via a use-after-free ($25,000); Compass Security took Cursor for an additional $15,000. Day 3 (ZDI, 2026-05-16; Hackread, 2026-05-16): STARLabs SG's Nguyen Hoang Thach burned a memory-corruption vulnerability for a full VMware ESXi hypervisor escape ($200,000, 20 Master-of-Pwn points); Windows 11 LPE chains landed; Compass Security attempted Claude Code but collided with a vulnerability ZDI already had on file. Master of Pwn final: DEVCORE 50.5 points / $505,000; STARLabs SG 25 points. Across three days: 47 unique zero-days, $1,298,250 paid out — ZDI's largest Berlin payout to date.

Exchange — compounding the in-the-wild picture. The DEVCORE three-bug chain attacks a different surface from CVE-2026-42897 (yesterday's deep dive) — OWA stored XSS is browser-context exploitation against authenticated users; the DEVCORE chain achieves SYSTEM-level direct RCE without authentication. Technique-class map: T1190 Exploit Public-Facing Application → T1059.003 Windows Command Shell → T1068 Exploitation for Privilege Escalation, with the EWS / RPC / RemotePS attack surface as the most plausible target set given Orange Tsai's prior ProxyLogon / ProxyShell / ProxyNotShell work. Embargo window: ZDI rules require vendors to ship patches within 90 days; expect Microsoft advisories around 2026-08-14, possibly bundled into August Patch Tuesday. Operational implication for the next ~12 weeks: on-premises Exchange faces (a) the currently-exploited XSS without permanent patch, (b) an unpatched unauthenticated SYSTEM RCE class proven viable on hardened production builds, and (c) the residual ProxyShell/NotShell attack surface that the FamousSparrow Azerbaijani campaign covered in the 2026-05-14 deep dive showed is still being weaponised against unpatched installs. The defender posture published with the 2026-05-16 deep dive (verify EEMS service, monitor OWA access patterns, restrict ECP/EWS from the internet, accelerate Exchange Online migration where possible) becomes harder to argue against given the Pwn2Own evidence.

AI Agents category — the new public-vulnerability surface for dev toolchains. Pwn2Own Berlin 2026 was the first year ZDI ran an AI Agents track. The result across the AI-Agents and adjacent inference-stack targets — OpenAI Codex, Cursor, LM Studio, LiteLLM, Claude Code, Claude Desktop, Chroma, Megatron Bridge, Ollama — was that the entries either landed exploits or collided with bugs ZDI already had on file (the latter still confirms the vuln exists). The recurring pattern across LiteLLM, LM Studio, Cursor and the OpenAI Codex attempts is agent-instruction-injection → server-side request forgery → arbitrary code execution, mapped to T1059.007 (JavaScript / scripting) and T1090 (Proxy abuse) — the agent runtime takes adversary-supplied content (a tool invocation, a file the agent is asked to summarise, a URL), treats it as a privileged instruction, and either fetches an attacker-controlled resource SSRF-style from inside the corporate network or executes attacker-shaped code in the agent's runtime container. STARLabs SG's five-bug LM Studio chain (Day 1) and k3vg3n's LiteLLM chain (Day 1) both follow exactly that pattern — SSRF→code-injection. OtterSec's Day 2 LM Studio pop was a code-injection bug only (no SSRF prefix), demonstrating the same target falls to two distinct attack-class roots. The OpenAI Codex CWE-150 vulnerability Compass Security exploited centres on improper neutralisation of special characters in tool-invocation arguments. Defender concepts that translate without IOCs: (1) treat self-hosted inference services (Ollama, LM Studio, LiteLLM, vLLM gateways) as untrusted public-facing applications even when bound to localhost — they are reachable from any browser tab the developer opens; (2) constrain outbound egress from inference containers to only the model-update endpoints they need (RFC-1918-range alerts from agent containers are a high-signal SSRF indicator); (3) require code-signing on tool plugins loaded by Cursor / Codex / Claude Code; (4) inventory developer endpoints that have agent tooling installed and ensure EDR coverage extends to the agent's runtime processes — these are not yet routinely covered by SOC tooling baselines. For Swiss/EU public-sector environments specifically: agentic coding tools are entering federal and cantonal developer workflows ahead of any procurement-grade evaluation; the Pwn2Own results give a documented evidence base for SOC managers asking developer-tooling owners for inventory and egress controls.

Capacity-overflow rejected-researcher PoC wave. A distinctive feature of Berlin 2026: Pwn2Own contest slots filled before all submitted research could be staged. ZDI's response — public disclosure of full PoC chains by researchers whose submissions were rejected for slot reasons — produced an unprecedented PoC release wave covering Firefox full-chain RCE, additional Ollama / LM Studio exploitation, NVIDIA driver chains, and at least one researcher's Claude Code exploitation attempt. Operationally, defenders cannot rely on the standard Pwn2Own embargo for any of these — the technical detail is in the wild now. Browser/inference/dev-tool teams should monitor researcher Twitter/Mastodon disclosure channels and triage against their own deployment surface immediately rather than waiting for vendor advisories.

Hardening / mitigation summary (citing the contest blogs for each piece):

  • Exchange on-premises: treat as severely threatened; verify EEMS M2.1.x; restrict OWA/ECP/EWS internet reachability; plan for an August Patch Tuesday emergency cycle when the DEVCORE embargo expires (ZDI Day 2).
  • VMware ESXi: assume the hypervisor escape class is exploitable on hardened production builds until Broadcom ships a patch; restrict ESXi management network reach; monitor for atypical guest-to-host process spawn patterns (ZDI Day 3).
  • AI Agents (Codex / Cursor / LM Studio / LiteLLM / Claude Code): treat inference containers as untrusted; egress-restrict to model-update endpoints; require tool-plugin code signing; inventory developer endpoints with agent tooling and ensure EDR coverage of agent runtime processes (ZDI Day 1, ZDI Day 2).
  • Windows 11 LPE candidates: track Patch Tuesday cadence ahead of August disclosure window; nothing actionable until vendors ship advisories (ZDI Day 3).

6. Action Items

  • Patch FunnelKit Funnel Builder for WooCommerce to v3.15.0.3+ immediately on any operator-managed WordPress instance and manually purge Settings > Checkout > External Scripts. Active exploitation across 40,000+ stores via unauthenticated checkout-endpoint injection; the WebSocket-fed skimmer is polymorphic per victim and will not be caught by static IOC matches. Reference: § 1 FunnelKit item.

  • Audit Polish-Qualified-Electronic-Signature workflows for SzafirHost 1.2.1 deployment if your environment accepts eIDAS-cross-recognised signatures from KIR Szafir signers (federal procurement, court e-filing counterparties, healthcare partners). The pre-patch zip-polyglot JAR bypass means SzafirHost endpoints exposed to a compromised download path could have produced silently-fraudulent signatures during the disclosure window. Reference: § 1 SzafirHost item.

  • Apply the F5 BIG-IP / BIG-IQ May 2026 Quarterly Notification across all internet-or-management-network-reachable F5 appliances within standard change windows, and rotate every Manager-role iControl REST credential at the same time. CVE-2026-41225 is post-auth Manager-role only, but the practical attack chain is credential theft → iControl REST object creation → shell execution, so credential rotation is the operationally critical companion to the patch. Reference: § 2 F5 item.

  • Enumerate and isolate any DHTMLX PDF Export Module endpoints (Node.js service backing DHTMLX Gantt / Scheduler / Diagram exports in EU public-sector portals); patch PDF Export Module to 0.7.6 and Diagram to 1.1.1. CVE-2026-41553 is unauthenticated, CVSS 4.0 score 10.0 — assume opportunistic scanning starts within days of CERT-PL's coordinated disclosure. If the patch cannot land within 48 h, restrict the export service to internal trusted origins via network ACL and apply egress filtering on the Node.js worker process. Reference: § 2 DHTMLX item.

  • Treat on-premises Microsoft Exchange as severely threatened through August Patch Tuesday 2026. Verify EEMS service is enabled and mitigation M2.1.x is in the applied list (Get-ExchangeDiagnosticInfo); restrict ECP / EWS / OWA reachability from the internet at WAF or reverse proxy where business-feasible; accelerate any in-progress Exchange Online migration. The compound risk (CVE-2026-42897 active XSS, no permanent patch + DEVCORE Pwn2Own SYSTEM RCE chain under 90-day embargo) does not have a single mitigation. Reference: § 4 UPDATE on Exchange.

  • Inventory AI agent / inference deployments on developer endpoints (LM Studio, Ollama, LiteLLM, Cursor, OpenAI Codex, Claude Code, vLLM gateways) and apply network egress controls so the agent runtime cannot reach RFC-1918 internal ranges. Pwn2Own Berlin 2026 demonstrated T1190 → T1090 → T1059 chains on every AI Agents target; SSRF-pivot to internal endpoints is the load-bearing technique. Require tool-plugin code signing where supported; ensure EDR coverage extends to the agent runtime processes (node, python, electron child processes spawned by agent UIs). Reference: § 5 deep dive — AI Agents section.

  • Hunt for Kimsuky's TryCloudflare-tunnel C2 pattern on government / defence / healthcare endpoints. Alert on outbound *.trycloudflare.com connections that don't correlate with a developer's legitimate Cloudflare-tunnel usage profile; review VSCode tunnel authentications via GitHub for unrecognised tunnel names; flag Rust-compiled PE images loading from non-standard paths. The German targeting documented in Kaspersky's report is the closest geographic-proximity Kimsuky signal to Swiss government estate in recent reporting. Reference: § 3 Kimsuky item.

7. Verification Notes

  • Items dropped:
    • GitLab CE/EE 18.11.3 security release — 25 CVEs including CVE-2026-7481 / CVE-2026-7377 / CVE-2026-6073 (CVSS 8.7 stored XSS). Dropped from § 2 because no CVE in the cluster clears any § 2 inclusion gate: not CISA-KEV-listed, no ENISA-EUVD exploited=true entry, top CVSS 8.7 (< 9.0), no public PoC, no in-the-wild exploitation reported. The XSS cluster is post-auth (Developer role required) with victim interaction prerequisite. Recommend operators apply through normal GitLab patch cadence; flagging would be over-weighted given today's competing operational signal. Source for reference: NCSC-NL NCSC-2026-0161, 2026-05-15.
    • FamousSparrow (UAT-9244) Azerbaijani oil & gas three-wave Exchange intrusion (Bitdefender, 2026-05-13). Surfaced by S3 and S4 but the primary publication is from 2026-05-13 — outside both the 40 h recency window and the 72 h developing window. The 2026-05-14 daily deep dive already covered FamousSparrow against the Azerbaijani energy operator; no material new development this run. Source for reference: Bitdefender Business Insights, 2026-05-13.
  • Single-source items: § 3 Kimsuky / PebbleDash / HelloDoor (Kaspersky Securelist as sole source) — included under the HIGH-reliability research-lab carve-out; explicitly flagged [SINGLE-SOURCE] in the item.
  • Items included with reduced confidence (window edge): § 3 Kimsuky item is at the edge of the 72 h developing window (Securelist publication 2026-05-14, ~62 h before run start) — included because the technical novelty (Rust-based PebbleDash variant + TryCloudflare quick-tunnel C2) is defender-actionable and not previously covered.
  • Contradictions: none surfaced this run.
  • Sub-agent timing: all four cti-research sub-agents (S1, S2, S3, S4) returned within budget. S1 → S4 durations 489 / 470 / 678 / 516 seconds, all well under the 30-min hard cap. S3 wrote S3.ended_at but its findings.S3.yaml was finalised at 04:19:21 UTC, with the assistant-text return delivered ~2 minutes after the disk artefact landed (no operational impact — main agent waits on the disk artefact, not the assistant-text return).
  • Coverage gaps: databreaches-net (5 consecutive runs failing 403 — Cloudflare anti-bot, no Wayback corroboration available; covered via primary regulator notices and victim disclosures where present); inside-it-ch (5 runs 403 — Cloudflare-protected, bridge fetches blocked); cisco-psirt RSS endpoint returned 404 in this run (last known good 2026-05-15; main vendor PSIRT page reachable, advisory list not); cert-fr (no new items since 2026-05-13 in the AVI/ACT feeds); cert-eu (most recent advisory dated 2026-05-13, outside the 40 h window); talos (bridge known-403, not attempted in this run — pivoted to other channels; pre-window content only); sophos-xops (HTTP 503 on blog fetch in this run); dfirreport (EtherRat + TukTuk flash alert published 2026-05-11, outside 72 h developing window). One candidate source surfaced by S3 — sansec-research (Sansec, sansec.io — Magecart and e-commerce skimming primary research) — promoted to status: "candidate" in sources/sources.json per the one-candidate-per-run rule.
  • Verification: five Phase 5.7 iterations ran (Opus / Sonnet-alt / Opus / Sonnet-alt / Opus rotation per v2.47); the brief publishes at the v2.46 cap-breach safety valve with the final iteration NEEDS_FIXES (truth=4, editorial=0). Iter1 found 5 truth + 1 editorial + 1 advisory (Pwn2Own attribution drift, F5 source-order, F5 quantifier); iter2 found 3 truth + 1 advisory (Satoki Tsuji attribution, CVSS dual-scale confusion 8.6 vs 9.1, OtterSec LM Studio technique over-claim); iter3 found 4 truth + 2 editorial + 1 advisory (SzafirHost overreach on Polish public-sector specifics, Swiss procurement portal claim, "dominant" quantifier, SecurityWeek phrasing, NCSC-NL CSAF source-attribution gap, Day 1 Orange Tsai Edge omission); iter4 found 1 truth + 1 editorial (NCSC-NL SPA URLs swapped for canonical CSAF JSON URLs, CWE-250 → CWE-648 per NVD). Iter5 residuals retained as the cap-breach record: F1 quote re-attributed (Sansec → The Hacker News citing Sansec, applied); F4 DHTMLX src qualifier scoped to CVE-2026-7182 only (applied). Iter5 findings F2 and F3 were verifier false-positives — main agent re-verified the NCSC-NL CSAF JSON directly (tools/fetch_source.py ncsc-nl csaf NCSC-2026-0162) showing 43 unique CVEs in the BIG-IP / BIG-IQ scope (not 23 as iter5 claimed) and all of CVE-2026-42930, CVE-2026-42924, CVE-2026-42406, CVE-2026-41953 present in the CSAF (iter5 claimed CVE-2026-42406 and CVE-2026-41953 absent). The brief preserves the 43-count and the four-CVE injection cluster as written. verification_residual_count = 4 records iter5's raw counts before this rebuttal-by-direct-CSAF-re-verification.