Security

AI agents wrote working exploits in under 4 hours and compromised 395 organizations in 48 countries

One operator's AI agents needed under four hours to reach a first real victim and compromised 395 organizations across 48 countries. A 23-year dataset from Andreessen Horowitz shows attack time has gone negative while fix times lengthen, and the bill is landing on corporate IT budgets.

Vincent JiangVincent Jiangfour-minute read
Share
Richard Horne, head of Britain's National Cyber Security Centre, seated on a cybersecurity panel at London Tech Week
1 / 6Slide 1 of 6
Richard Horne, head of Britain's National Cyber Security Centre, on a London Tech Week panel on harnessing AI for cybersecurity, June 2025.

One print server, a few hundred agents

On August 31 a single operator, likely Russian-speaking, pointed a few hundred AI agents at PaperCut NG and MF, the print-management software used by schools and offices worldwide 13. Working from an empty workspace, the agents wrote exploits for two bugs disclosed four days earlier and reached a first break-in at a real victim in under four hours; 11 organizations fell in 26 seconds once the campaign launched 13.

395 organizations, half of them schools

GreyNoise, the threat-intelligence firm that tracked the campaign, counted 440 compromised servers at 395 organizations in 48 countries 1, half of them schools 3. The agents ran on a cheap open-weight model, DeepSeek, inside OpenAI's$1.18T — OpenAI, private, latest valuation $1.18T Codex agent harness, and ignored their operator's list of 28 countries to avoid 123. Andreessen Horowitz (a16z), the venture capital firm, published a 23-year dataset on the gap between attack speed and patch speed today, and its conclusion lands directly on corporate IT budgets: the cost of breaking in collapsed while the cost of fixing did not move 3.

The 22-day expert became a $4,000 line item

For twenty years the bottleneck in breaking in was the expert who turns a bug into a working exploit; RAND's 2017 study of more than 200 zero-days priced that job at a median 22 days of professional work 3. Anthropic$350B — Anthropic, private, latest valuation $350B, the frontier lab, reported in March 2026 that about $4,000 and several hundred attempts bought two working Firefox exploits from its then-current model, and that its next model, kept off general release, produced 181 on the same task 3.

Attack time went negative; fix time went the other way

Mandiant, Google's threat-intelligence unit, measured the average time from disclosure to first attack at 63 days in 2018-19, five in 2023, then across the zero line: exploitation began one day before disclosure in 2024 and seven days before in 2025 3. Fixing moved backward. Verizon's−0.44% — Verizon, down 0.44 percent as of its last quote 2026 Data Breach Investigations Report puts the median time to fully fix a known-exploited vulnerability at 43 days, up from 32 a year earlier, with only 26 percent fully fixed within the year 3.

The average attack now starts before the patch exists

Exploitation crossed zero: attacks now begin before disclosure

-20 days0 days20 days40 days60 days80 days2018-192020-212021-22202320242025Disclosure dayAttacks now start 7 days beforethe patch
Data
Disclosure to first attack
2018-1963 days
2020-2144 days
2021-2232 days
20235 days
2024-1 days
2025-7 days
Average days from vulnerability disclosure to first observed attack; Mandiant incident-response data as compiled by a16z on November 5, 2026. Values below zero mean exploitation began before disclosure. Source: a16z citing Mandiant.3

Black October priced the gap

Binance, the crypto exchange, now puts its October 29 theft at $6.34 billion, reached through malware planted during a fake job interview on one engineer's laptop, with no new vulnerability involved; about $2.85 billion of the stolen assets has since been frozen 34.

Richard Horne, head of Britain's National Cyber Security Centre, warned on November 3 that vulnerabilities "tolerated for far too long" can now be exploited "at a scale that was once unimaginable" 5. Analysts call the stretch "Black October" 6.

When breaking in costs hours of machine time, nobody is too small to target: the exposed base includes 2.6 to 4.4 million internet-facing installations of PHP, the programming language behind most websites, about 80 percent of them on versions no longer supported 36.

The skeptics sit inside the same firm

The other camp inside a16z recalls that Metasploit, the 2003 toolkit that armed amateurs, and Shodan, the 2009 search engine for internet-exposed devices, were each supposed to end the internet, and neither did 3. Ransom payments have fallen for two straight years even as leak-site claims rose, and the PHP bugs produced few confirmed compromises once hosting providers blocked attacks at their edge within a day 3. The firm's own bet is a long, expensive worm era for anyone without a platform to shelter under, not a blackout 3.

Budgets shift to platforms that block attacks at the edge

Big institutions are rerouting up to a fifth of IT resources to finding and fixing flaws, says Vanessa Lyon of Boston Consulting Group, and Goldman Sachs−0.28% — Goldman Sachs, down 0.28 percent as of its last quote has lifted the security share of its AI budget 6.

During the PHP incident, edge firewalls from Amazon−1.19% — Amazon, down 1.19 percent today and Cloudflare−4.73% — Cloudflare, down 4.73 percent today covered small sites that could not patch themselves 6. Amazon's chief information security officer, CJ Moses, whose MadPot decoy servers watched one attacker go from no access to full system takeover in 12 minutes 42 seconds, says patch service-level agreements of 7 to 30 days now "assume a delay that no longer exists" 37.

CJ Moses, chief information security officer of AWS
CJ Moses, Amazon's chief information security officer, whose MadPot decoy servers track AI-era attackers.Courtesy of Amazon

2,000 exploited bugs by 2027 is the test

VulnCheck, which tracks which bugs are exploited in the wild, counted about 880 exploited vulnerabilities in 2025; if the count clears 2,000 by 2027, the cost collapse has arrived in full 3.

Get Security by email, free

The breaches, flaws and security vendors AI is putting to the test. In our daily email, every weekday.

Loading VZ

Reader comments

0 comments

    More in Security

    All in this topic →

    Security

    Okta's AgentCore deal with Amazon has no disclosed price, and its fiscal 2027 guide carries no agent revenue

    AWS's November 12 AgentCore update ships every session with its own identity, policy and audit trail, the same guarantee Okta sells, and Okta's filed fiscal 2027 guide carries no agent revenue. OKTA is up about 106 percent this year into a December 1 print that has to show the money.

    Security

    Palo Alto's Unit 42: Chinese attackers ran half of 25 autonomous AI hacks on Zhipu's GLM 5.2

    Unit 42's telemetry shows one operator used Zhipu's GLM 5.2 for 13 of 25 autonomous intrusions that reached a teachers' union and .mil and .gov accounts. The report hands Washington a named Chinese model as chip curbs gain votes, and lands on Z.ai's Hong Kong holders.

    Security

    SentinelOne signs $400 million credit line secured by substantially all its assets

    The five-year Bank of America line costs roughly $1.0 million to $1.4 million a year undrawn, and if it is ever drawn its lenders rank ahead of October's $1.25 billion of convertible noteholders in any claim on the company's assets. The next clue lands with earnings on December 3.

    Security

    Cellebrite beats Q3 and lifts EBITDA guidance as CFIUS weighs divesting Corellium

    Revenue, ARR and free cash flow all beat targets, but the same US government buying faster than ever is still deciding whether Cellebrite must divest Corellium, the $170 million acquisition it closed without formal clearance.

    Security

    France blames Iran's Revolutionary Guards for 700-store sabotage and says they probed Schneider Electric gear

    The same French report says the attackers' infrastructure was probing Schneider Electric equipment at energy and water sites. For holders of the biggest listed industrial-control vendor, that is event risk on the installed base, and the week's price action does not reflect it.

    Security

    MongoDB patched a 9.5-severity flaw that crashes servers at failover; self-managed users must upgrade

    Atlas patches itself during maintenance, while self-managed customers must upgrade every replica-set member by hand across seven branches. That points the cost at MongoDB's legacy license base rather than the cloud growth behind BTIG's $417 Buy, ahead of the December 1 earnings.

    Subscribe to TheInference for free