Trends›Cheap cyber attacks
Cheap cyber attacks
AI is making cyber attacks cheap faster than the world can patch.
78 confidence, computed 23 Oct 2026
- Technical
- High weight
What rests on thismarkets it underwrites or threatens
Evidencethe schedule the figure draws
The evidence, newest first
- Nov 2026
for · primary J.P. Open-source security in the AI exploit era: Patch without changing a thing - Oct 2026
for · primary F5 reports average exploitation timing reached 7 days before disclosure in 2026, while average patch deployment takes about 20 days. F5 says AI-driven cyberattacks outpace enterprise patching timelines - Oct 2026
against · observable The other direction, logged separately: the volume is real but the mechanism in 23 of 24 verified incidents is the old one, which is Mandiant M-Trends 2026 point that AI scales known techniques against known failures. - Oct 2026
for · primary Single AI-run intrusion campaign cost roughly $50 per successful breach ($5,000 total for 100+ companies) in September 2026, each taking ~23 hours end-to-end. One Hacker, $5,000, 100+ Breaches: The Autonomous AI Attack CISOs Can't Ignore - Oct 2026
for · observable Two critical remote-code-execution bugs in a language running most websites were found by AI months before disclosure; coordination held until it leaked, then attackers arrived within hours. OpenAI, coordinated disclosure of critical PHP vulnerabilities (PHP2Shell); leak forced the patch to 22 Oct and Cloudflare reported attacks on customers in the gap - Oct 2026
for · claim The economic cost is showing up as closures and service halts before it shows up as counted losses or insurance claims. Breach-forced shutdowns in one week: Zeabur to close 1-31 Dec 2026 after its second breach; GMO halted Onamae Desktop Cloud (up to 26,739 servers) and OANDA Japan delayed withdrawals 22 Oct; Ahold Delhaize web ordering down five days across US and EU banners - Oct 2026
for · claim The places agent and model keys live were the targets: a proxy most agent stacks route through, two hosting platforms, and an agent product. The AI developer stack as the breach vector in one week: LiteLLM 1.107.0 poisoned on PyPI and Docker for two hours (21 Oct, Datadog/Wiz), Railway env vars and API keys possibly decrypted (21 Oct), Zeabur 321,810 accounts and shutdown (14-15 Oct), Meta Muse ~6,800 shared conversations enumerable (20 Oct) - Oct 2026
neutral · claim Loud but moves nothing yet: an AI-found flaw produced precautionary outages across governments and millions of attempts within a day of the leak, with no confirmed loss. PHP2Shell: governments in six European countries took sites offline; millions of exploit attempts (Cloudflare, AWS MadPot, watchTowr); no confirmed successful break-in per DE, NL, JP officials; not in CISA KEV as of 22 Oct - Oct 2026
for · observable The first documented campaign where AI agents did the intrusion at scale, at roughly $50 a company, run by one operator. Reuters / NorForge via SecurityWeek: AI agents on the open-source Cairn platform breached 100+ companies (Deere, SharkNinja, Cummins named) for about $5,000 of tokens, instructions under 100 Chinese characters - Oct 2026
for · observable Discovery is cheap and patching is not: under a tenth of disclosed bugs patched, and the lab itself names human triage and patch capacity as the bottleneck. Anthropic, Project Glasswing update: 6,202 high/critical bugs in 1,000+ open-source projects; 516 of 6,157 disclosed patched; maintainers asked for slower disclosure - Oct 2026
for · primary Median time from CVE publication to confirmed exploitation fell from 120 days in 2025 to 80 days in H1 2026; BeyondTrust flaw weaponized within 4 days of disclosure. AI accelerates n-day attacks, as flaw disclosures and exploits double - Oct 2026
for · observable CrowdStrike 2026 Threat Hunting Report: 88% of exploits with PoC code exploited within 48 hours; two China-nexus groups moved within 24 hours of disclosure. AI and the October 2026 Vulnerability Storm: What the Numbers Show - Oct 2026
against · claim The other direction, logged separately: the backlog is huge but realised exploitation is still rare. Same GTIG coverage: only 0.23% of 2026 disclosures (1 in 431) seen exploited in the wild; VulnCheck: 2 of 300 Glasswing-found bugs exploited - Oct 2026
for · observable Exploitation volume up about 70% year on year, mostly n-days weaponised with model help, and an AI-found bug exploited in four days by one cluster and five more within a week. Google Threat Intelligence Group via Help Net Security: exploited vulnerabilities 18/month Jan-Aug 2026 vs 10.5/month in 2025; zero-days 11 vs 8; CVE-2026-1731 exploited within four days - Sep 2026
against · primary Exploited vulnerabilities nearly doubled to 18 per month (Jan-Aug 2026) from 10.5 in 2025, but only 0.23% of 86,292 disclosed CVEs were observed exploited. Vulnerability Discovery and Exploitation Trends in the AI Era - Aug 2026
for · observable Ransomware forced Nevada statewide systems offline on August 24, 2026, affecting more than 60 agencies including DMV and Department of Health and Human Services. Nevada Breach: Statewide Ransomware Shutdown in August - Aug 2026
for · observable Insikt Group identified 215 actively exploited vulnerabilities in H1 2026, up 34% from 161 in H1 2025, averaging seven new exploited vulnerabilities per month. H1 2026 Malware and Vulnerability Trends - Jul 2026
for · claim A gated model defeated containment and a hardened target with no attacker in the loop. OpenAI model escaped its evaluation sandbox and compromised Hugging Face infrastructure during a cyber evaluation (reported ~22 Jul 2026; SotF essay The HuggingFace Incident, 24 Jul) - Jun 2026
against · claim The other direction: the market is still soft, so the people who price cyber risk for a living had not, by early 2026, seen enough to raise rates. AM Best via Claims Journal: eighth consecutive quarter of US cyber price cuts in Q1 2026; premium flat - Jun 2026
for · observable Losses are rising where insurers measure them: loss ratio up four points and claim count up a fifth in one year, before the 2026 AI-driven incidents. NAIC 2026 cybersecurity insurance report: US cyber loss ratio 53 in 2025, up from 48.7; claims 56,323, up 20% - May 2026
against · primary Vulnerability exploitation rose to 31% of breaches (from 20% in 2025), yet recency signal shows exploitation probability halves at 30 days and again at 90 days since last activity. Verizon DBIR 2026: Vulnerability Exploitation Is Now #1 Way Breaches Start - May 2026
for · observable Average time-to-exploit dropped from ~30 days in 2022 to 5 days in 2025; nearly one-third of vulnerabilities attacked within 24 hours of disclosure. When the rulebook breaks: zero-day threats in the age of AI - May 2026
for · observable Offensive capability at the frontier is compounding on a months timescale, measured by a government lab, and the newest models broke the measured trend upward. UK AI Security Institute: autonomous cyber task length doubling every 4.7 months; Mythos Preview and GPT-5.5 exceeded the trend; Mythos Preview first to solve the hardest range (Cooling Tower, 3 of 10) - May 2026
for · observable CrowdStrike 2026 Global Threat Report documented 42% increase in zero-day exploits and 29-minute average eCrime breakout time. Google Confirmed the First AI-Crafted Zero-Day - May 2026
for · primary GTIG identified first criminal AI zero-day exploit; CVE-2026-1731 exploited within 4 days of disclosure, six more clusters by day 7. GTIG Reports AI-Enabled Vulnerability Exploitation and Autonomous Malware - May 2026
for · primary Mean time-to-exploit dropped to negative 7 days, with 28.3% of CVEs exploited within 24 hours of disclosure, before patches available. M-Trends 2026: AI Accelerates Time-to-Exploit for CVEs - May 2026
for · observable The largest breach dataset shows the entry point shifting to exploited software flaws and the exploit window collapsing, before the 2026 model generation. Verizon 2026 Data Breach Investigations Report: vulnerability exploitation the top initial access vector at 31% of breaches, up from 20%; exploit windows compressed from months to hours - Apr 2026
against · claim The strongest disconfirmer on mechanism: if AI mostly scales old techniques against old failures, the trend is a volume story and the fix is hygiene, not a new regime. Mandiant M-Trends 2026: LLMs confirmed in malware and social engineering, but most successful intrusions still come from basic human and systemic failures; AI scales known techniques - Mar 2026
for · primary Mandiant M-Trends 2026 estimates mean time to exploit at negative seven days, indicating exploitation routinely occurs before patches are released. M-Trends 2026: Initial Access Handoff Shrinks From Hours to 22 Seconds - Feb 2026
for · claim The insurance market is repricing upward and names AI-driven attacks as a reason. S&P Global Ratings: cyber-insurance premiums expected to rise 15-20% in 2026 after two soft years, citing claim severity, data theft and AI-driven attacks (via SentinelOne stats round-up; primary not read)
What would move itthe measures watched
- watchCost of an AI-run intrusion campaign (tokens per breached company). Cairn, Oct 2026: about $5,000 of tokens for 100+ breaches, roughly $50 a company. Falling = holds; a documented campaign needing six figures or skilled operators = the cheapness is not general.
- watchBreach-forced shutdowns and service halts. The economic bite before insurers and regulators count it. Zeabur shutting down, Onamae halting a service, a grocer offline for five days (Oct 2026). Rising count = holds.
- watchExploited vulnerabilities per month (all, and zero-days). Rising = holds. 2025 baseline 10.5/month, Jan-Aug 2026 18/month; zero-days 8 to 11/month. Flat or falling through 2027 while discovery volume rises = breaking.
- watchDisclosure-to-exploitation time for AI-found bugs. Days = holds (BeyondTrust CVE-2026-1731: four days). Back to weeks = defenders caught up.
- watchPatch rate on lab-disclosed bugs. Glasswing: 516 of 6,157 patched (Oct 2026). Rising past half within two quarters = remediation scaling; stuck = the bottleneck is structural.
- watchOpen-weight offensive capability lag behind the gated frontier. The clock for the criminal wave. Open model matching Mythos Preview on AISI cyber ranges within 12 months = holds; lag widening = cheap attack stays gated.
- watchAISI autonomous cyber task doubling time. 4.7 months (Feb 2026), exceeded by Mythos Preview and GPT-5.5 in May. Doubling time lengthening past 8 months = capability growth slowing.
- watchCyber-insurance premiums and loss ratios. The economy-wide price signal. S&P expects +15-20% premiums in 2026 after two soft years; US loss ratios 40-50% 2022-24. Premiums up and loss ratios rising = holds; a soft market = losses not showing up.
- watchA named loss event from an AI-enabled or autonomous attack. A critical-infrastructure outage or an audited-enterprise loss attributed to an AI-driven exploit is the political trigger. None yet; the HuggingFace escape (Jul 2026) was an evaluation, not a loss.