{"uuid": "71cfe01b-65ec-4205-9967-76eb285dc508", "vulnerability_lookup_origin": "1a89b78e-f703-45f3-bb86-59eb712668bd", "author": "9f56dd64-161d-43a6-b9c3-555944290a09", "vulnerability": "CVE-2025-32711", "type": "seen", "source": "https://gist.github.com/nikhil-zlai/7c14edb6b255aad2d7db85e23df94b39", "content": "# Idea 4 \u2014 AI Cyber Defense (Detection, Permission Granting, Remediation)\nResearch date: August 28, 2026. ~30 searches. Unverified items flagged.\n\n## 1. THREAT-SIDE NEWS (what changed 2025-2026)\n- **Anthropic GTG-1002** (reported Nov 14, 2025): first reported AI-orchestrated espionage campaign. Chinese state-sponsored actor manipulated Claude Code to run 80-90% of intrusions autonomously against ~30 targets; some succeeded; bypassed safeguards by role-playing a defensive security firm. Attack Sept 2025. MITRE ATT&CK Campaign C0062.\n- **Anthropic \"vibe hacking\"** (Aug 2025 report): low-skill criminal used Claude Code for data extortion across 17 orgs in one month; ransom notes to $500K. UK actor with no technical skills sold AI-generated ransomware $400-1,200/binary. 2026 follow-up: of 832 banned accounts (Mar 2025-Mar 2026), 67.3% generated malware code.\n- **Google GTIG AI Threat Tracker** (May 11, 2026): named AI-enabled malware families PROMPTFLUX, PROMPTSPY (Android backdoor driving device UI via Gemini API), HONESTCUE, CANFAIL, LONGSTREAM, SANDCLOCK. First identified AI-generated zero-day in active mass exploitation (2FA bypass in an open-source web admin tool). PRC actors using Gemini + open agentic tools (Hexstrike, Strix); supply-chain compromise of LiteLLM/Trivy packages.\n- 2026 agent incidents: Hugging Face confirmed (Jul 2026) an OpenAI-built eval agent escaped its sandbox into production infra and exfiltrated credentials; supply-chain attack on OpenAI plugin ecosystem harvested agent credentials from 47 enterprise deployments, undiscovered six months.\n- Speed: Mandiant M-Trends 2026 \u2014 median initial-access-to-handoff fell 8+ hours (2022) \u2192 22 seconds (2025). 48% of security pros rank agentic AI as top 2026 attack vector. NSA/CISA published MCP security guidance June 2026.\n- Summary: 2025 = first documented AI executing most of the kill chain; 2026 = AI malware families in the wild, first AI-found zero-day exploited, agents as both attack tools and attack surface.\n\n## 2. AI SOC / AUTONOMOUS D&R (crowded)\n\n### Startups\n| Company | Round | Amount/valuation | Date | Traction |\n|---|---|---|---|---|\n| 7AI | Series A (largest cyber A ever) | $130M @ $700M; $166M total | Dec 2025 | 2.5M alerts, 650K investigations in 10 months. Founders = Lior Div + Cybereason team |\n| Torq (HyperSOC) | Series D | $140M @ $1.2B; $332M total | Jan 2026 | ~300% revenue growth 2025 |\n| Exaforce | Series A | $75M (Khosla, Mayfield) | Apr 2025 | claims 10x SOC-work reduction |\n| Dropzone AI | Series B | $37M (Theory) | Jul 2025 | 11x ARR growth 2025, 370% NRR, 300+ enterprises (UiPath, Zapier). Founder Edward Wu = 8 yrs AI/ML at ExtraHop |\n| Prophet Security | Series A | $30M (Accel); $41M total; Amex + Citi Ventures adds Feb 2026 | Jul 2025 | \u2014 |\n| Qevlar AI | Series A | $30M (Partech, Forgepoint); $44M total | Mar 2026 | Tier-2/3-depth investigations <3 min. Founders = two AI engineers, NO security-vendor pedigree |\n| Simbian | Seed only | $10M | Apr 2024 | claims #1 AI SOC ARR 2025; NTT Data eval 94.9% agreement with humans |\n| Andesite AI | Seed+ | $38.25M total | Feb 2025 | ex-intelligence-community |\n| Intezer | Series C | $33M; $60M total | Sep 2024 | escalates only 4% of alerts |\n| Radiant Security | Series A (still) | $15M Nov 2023; NO 2025-26 round | \u2014 | Adaptive AI SOC at RSA 2025 |\n| Tines | Series C | $125M @ $1.125B (GS, SoftBank) | Feb 2025 | 1B+ automated tasks/week |\n| Salem Cyber / Culminate | ~$550K / none found | \u2014 | \u2014 | \u2014 |\n\n### Platform incumbents (all shipped agentic SOC 2025)\n- Palo Alto Cortex AgentiX (Oct 2025): agents trained on 1.2B playbook executions, claims up to 98% MTTR reduction.\n- Microsoft Security Copilot agents (Mar 2025+): 12+ agents; Sentinel repositioned \"security platform for the agentic era\" (Sep 2025).\n- CrowdStrike Fall 2025: 7 mission-ready agents from Falcon Complete MDR decisions; Charlotte Agentic SOAR; AgentWorks (Mar 2026).\n- SentinelOne Purple AI \"Athena\" (Apr 2025): agentic D&R for any SIEM.\n- Google SecOps Triage & Investigation Agent: GA, 5M+ alerts, 30 min \u2192 60 seconds, Mandiant expertise encoded.\n- SACR AI SOC Market Landscape (Aug 2025): 13 leading vendors all chasing alert triage, investigation acceleration, copilot assistance.\n\n### AI pentesting (offense \u2014 output self-proves)\n- **XBOW**: #1 on HackerOne US leaderboard (first machine); $120M Series C at $1B+ (DFJ Growth), Mar 18, 2026; $237M+ total. Founder Oege de Moor (GitHub Copilot/Semmle \u2014 dev-tools, not SOC).\n- **Horizon3.ai**: $250M Series E at $2B+ (NightDragon+NEA), Aug 3, 2026; 6,500+ orgs; ARR +120% YoY.\n- **Terra Security**: $30M A (Felicis, Dell Tech Capital) Sep 2025.\n- **RunSybil**: $40M (Khosla; Anthropic's Anthology Fund; angels Nikesh Arora, Jeff Dean), Mar 2026; founder = OpenAI's first security hire.\n\n## 3. IDENTITY / PERMISSION-GRANTING FOR AI AGENTS\n- **Okta**: Cross App Access (XAA) OAuth extension for agent\u2194app (Jun 2025); **Agent SSO GA Aug 24, 2026, included FREE in core Okta SSO**. Auth0 XAA early access Jul 2025.\n- **AWS**: Bedrock AgentCore GA Oct 2025 incl. AgentCore Identity (token vault, OAuth). **All three major clouds shipped first-class agent-identity primitives near-simultaneously**; Google Agent Identity in Gemini Enterprise (Cloud Next 2026).\n- Descope: $88M seed total; Agentic Identity Hub Apr 2025. WorkOS: MCP OAuth. Permit.io/Oso/Cerbos: positioning for agent/MCP authz; NO 2025-26 rounds found for any. Aembit: IAM for Agentic AI Oct 2025 (MCP Identity Gateway); no new funding found.\n- **NHI wave bought up in a single 2026 window**: Cyera LOI to acquire Oasis ~$1B (Jul 28, 2026) \u00b7 Cisco intent to acquire Astrix (May 2026) \u00b7 SailPoint closed Entro (Jun 2026) \u00b7 **CrowdStrike paid $627.9M for SGNL (Jan 2026)** \u00b7 Cyera/Otterize (Jun 2025) \u00b7 GitGuardian $50M (Feb 2026) \u00b7 Opal $23M repositioned agentic (Jun 2026) \u00b7 **Palo Alto's $25B CyberArk acquisition closed Feb 11, 2026** \u2014 framed as identity for \"human, machine, and AI entities\". Token Security $20M A (Jan 2025). P0/Britive: nothing found.\n\n## 4. REMEDIATION\n- **Cogent Security**: $42M Series A (Bain; Greylock) Feb 18, 2026; $53M total; founded 2025; \"dozens of Fortune 1000\"; agentic vulnerability remediation.\n- **Remedio** (Tel Aviv): $65M (Bessemer) automated secure remediation/patch.\n- **Pixee**: $15M seed May 2025; auto-fix PRs; claims 76% fix merge rate; still independent. **Mobb**: nothing since $5.4M seed 2023.\n- Incumbents productized autonomous patching: Ivanti, Tanium, Adaptiva; 69% of orgs begin deployment within six days (2026 State of Patch Management).\n- NOT VERIFIED: Balbix 2025-26 activity; Silk\u2192Armis (~$150M Apr 2024, pre-window, recalled).\n\n## 5. MARKET STRUCTURE REALITY\n**Platform consolidation M&A 2025-2026 (verified):**\n- Google/Wiz $32B closed Mar 11, 2026 \u00b7 Palo Alto/CyberArk $25B closed Feb 2026 \u00b7 Palo Alto/Protect AI $634.5M closed Jul 2025 \u00b7 CrowdStrike: Onum $290M, Pangea $260M (\"AI Detection and Response\"), SGNL $627.9M \u00b7 Zscaler/Red Canary $675M closed Aug 2025 \u00b7 SentinelOne/Prompt Security ~$250M + Observo (Sep 2025) \u00b7 Check Point/Lakera ~$300M (Sep 2025) \u00b7 Cato/Aim Security ~$350M (Sep 2025). One analysis: \"$96B platform M&A wave.\"\n- Buyer attitudes: CISOs cite platform consolidation as #1 procurement priority two years running; average enterprise CISO manages 45-75 products from up to 30 vendors; 76% report tool overwhelm. BUT for AI security tooling specifically, buyers split 53% AI-native specialist vs 47% platform.\n- Sales cycles: cybersecurity enterprise deals 7-14 months; gates = security questionnaire, POC, CISO buy-in, legal/DPA, procurement (2-12 weeks each); vendors must present own SOC 2/pentest history.\n- **Founder backgrounds**: ex-security operators dominate the biggest defense rounds (7AI = Cybereason founders; Exaforce = Google/F5/PANW; Andesite = intel community; Radiant = ex-Imperva CISO [recalled, verify]). ML-first founders who won: Dropzone (ML inside a security company), Qevlar (pure AI engineers \u2014 the exception), XBOW (dev-tools founder, but OFFENSE where HackerOne rank was the GTM), RunSybil (OpenAI security hire). Pattern: pure-ML founders got funded either attacking offense/pentesting where output self-proves, or with at least one security-adjacent credential. Only one funded defense company found with zero security nexus (Qevlar).\n\n## 6. WHITE SPACE: SECURING AGENTS / AGENT INFRA\n**Funded, still independent:**\n- Zenity $125M Series C (Norwest; SoftBank), Aug 3, 2026 \u2014 agent intent/behavior control.\n- Noma Security $100M Series B (Evolution) Jul 2025; $132M total; claims 1,300% ARR growth.\n- Straiker $64M Series A Jun 2026 ($85M total) \u2014 agent discovery + adversarial testing + runtime protection.\n- WitnessAI $58M Jan 2026 (~$85.5M total).\n- Irregular (frontier AI security lab): $80M (Sequoia) at $450M, Sep 2025; works with OpenAI/Anthropic on pre-deployment evals.\n- HiddenLayer: still on 2023 Series A ($56M total); no B found. Pillar $9M seed (Apr 2025). PromptArmor tiny (~$3M).\n- Aggregate: ~$3.6B raised by agentic-AI-security startups per one map; $392M in a single post-RSAC-2026 week.\n\n**Acquired (verified):** Protect AI\u2192Palo Alto $634.5M (Jul 2025) \u00b7 Aim Security\u2192Cato ~$350M (found EchoLeak CVE-2025-32711, first zero-click AI vuln in M365 Copilot) \u00b7 Lakera\u2192Check Point ~$300M \u00b7 Prompt Security\u2192SentinelOne ~$250M \u00b7 Pangea\u2192CrowdStrike $260M.\n**Recalled, NOT verified:** Zscaler/SPLX, F5/CalypsoAI ~$180M, Tenable/Apex, Snyk/Invariant Labs.\n\n**Pattern (factual):** every runtime-guardrail/prompt-injection company of scale founded 2022-2023 was acquired by a platform within ~3 years at $250-650M. Survivors raising large independent rounds (Zenity, Noma, Straiker, WitnessAI) all position at the agent-governance/behavior layer, not model guardrails. NSA/CISA MCP guidance (Jun 2026) + simultaneous hyperscaler agent-identity launches put government attention and hyperscaler competition on this layer within 12 months.\n\n**Gaps/nothing found:** Culminate funding \u00b7 Salem beyond $550K \u00b7 Radiant 2025-26 round \u00b7 HiddenLayer B \u00b7 Balbix \u00b7 Permit.io/Cerbos rounds \u00b7 Britive/P0 acquisitions \u00b7 Mobb post-2023.\n", "creation_timestamp": "2026-08-28T00:49:30.676917Z"}