$1,279 divides total cost by 948 distinct patch-level LPEs. Dividing by the 1,169 working exploits gives about $1,038, and OpenAI says a successful exploit-dev attempt averaged about $360. None is a universal marginal cost of a new zero-day.
Back to the trend of the day: OpenAI says its AI agents found hundreds of ways to take over Linux
Poly Rating · from the trend of the day
AI in Vulnerability Research: Zero-Day Bounty Targets 2026
OpenAI's agent pipeline produced 948 distinct Linux root exploits for $1,213,430 in API credits, about $1,279 each, against $20,000–$70,000 historical contest prizes for the same bug class.
- No. 1: Linux Kernel
- 10 places
01Final list
The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.
Order: from most to least AI agreement.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Linux Kernel
$1,279 per distinct root exploit vs $20K–$70K contest prizes, a 16–55x gap. The $18.9M–$66.3M valuation is hypothetical, not actual payouts.
Included it: ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek. Did not include: Nemotron, Nova, MiMo, Wenxin, Qwen.In their lists: ChatGPT: 1 · Meta: 1 · Grok: 1 · Gemini: 1 · Kimi: 1 · DeepSeek: 1
What each AI noted 6
- ChatGPTplace 1
948 distinct privilege-escalation vulnerabilities demonstrated in one $1.21 million research campaign.
Source ↗ - Metaplace 1
948 working root exploits from one $1.21M run, about $1,279 per unique LPE against $20,000 to $70,000 in historical contest value.
Source ↗ - Grokplace 1
$1,279 AI cost vs $20k–$70k contest payouts yields 16–55× asymmetry on 948 distinct root paths
- Geminiplace 1
An autonomous agent pipeline extracted 948 root-level LPE exploits at $1,038 per weaponized bug against a target where human market bounties traditionally claim $70,000 to $100,000.
- Kimiplace 1
One 30-hour, $1.2M run minted 948 root exploits at $1,279 each — prices that humans only get paid $20,000-$70,000 for at Pwn2Own and TyphoonPWN.
Source ↗ - DeepSeekplace 1
$1,279 buys a root exploit that kernelCTF pays $71,337 for — a 56x gap on the same bug class.
Source ↗
6 AIs put it here - 2
Google Chrome
The one real payout: about $40,000 of AI compute bought a $250,000 full-chain award. The ratio is only 6.25x, and the cost excludes labor.
Included it: ChatGPT, Meta, Gemini, Kimi, DeepSeek. Did not include: Grok, Nemotron, Nova, MiMo, Wenxin, Qwen.In their lists: ChatGPT: 3 · Meta: 2 · Gemini: 6 · Kimi: 2 · DeepSeek: 2
What each AI noted 5
- ChatGPTplace 3
A $250,000 full-chain bonus coexists with much smaller rewards for individual memory-corruption findings.
Source ↗ - Metaplace 2
One researcher plus public GPT-5.6 Sol built a 3-bug full chain in 6 days for about $40,000 and took Google's first $250,000 full-chain reward.
- Geminiplace 6
Broker payouts reach $3,500,000 for browser RCE and sandbox escape chains that autonomous agent pipelines now reconstruct from commit logs for roughly $3,000.
- Kimiplace 2
A three-bug full Chrome chain that pays $250,000 was assembled by four people and AI agents in six days on ~$40,000 of compute — work that used to take a team 6-9 months.
- DeepSeekplace 2
$40,000 of AI compute earned a $250,000 Chrome full-chain bounty — the gap narrows when the target is a browser, not a kernel.
Source ↗
5 AIs put it here - 3
Apple iOS and macOS
The biggest vendor ceiling: $2M for a chain, over $5M with bonuses. No AI cost is published, and Apple capped AI-driven submissions.
Included it: ChatGPT, Meta, Gemini, Kimi, DeepSeek. Did not include: Grok, Nemotron, Nova, MiMo, Wenxin, Qwen.In their lists: ChatGPT: 4 · Meta: 6 · Gemini: 2 · Kimi: 4 · DeepSeek: 3
What each AI noted 5
- ChatGPTplace 4
Apple's largest publicly advertised exploit-chain rewards exceed $5 million with bonuses.
Source ↗ - Metaplace 6
Apple doubled its zero-click chain to $2M, with bonuses over $5M, matching broker money for the first time.
Source ↗ - Geminiplace 2
The $5,000,000 commercial zero-click bounty ceiling faces sub-$20,000 multi-agent reasoning chains that decompose sandbox escape primitives into automated AST fuzzing steps.
- Kimiplace 4
Apple raised its own ceiling to $2M per chain (above $5M with bonuses) just as agents learned to chain sandbox escapes elsewhere — the premium target is one public AI-cost datapoint away from repricing.
Source ↗ - DeepSeekplace 3
Apple's ceiling is $5M with bonuses, but AI found 50+ macOS flaws in three weeks — and Apple capped submissions.
Source ↗
5 AIs put it here - 4
Microsoft Windows and Hyper-V
Hyper-V pays up to $250,000 and Windows RCE $100,000. A 2019 broker list had $1M for Windows zero-click, and no AI cost is published.
Included it: ChatGPT, Meta, Gemini, Kimi, DeepSeek. Did not include: Grok, Nemotron, Nova, MiMo, Wenxin, Qwen.In their lists: ChatGPT: 2 · Meta: 7 · Gemini: 5 · Kimi: 6 · DeepSeek: 4
What each AI noted 5
- ChatGPTplace 2
Microsoft publicly offers up to $250,000 for a qualifying guest-to-host compromise.
Source ↗ - Metaplace 7
A broker paid $1M for zero-click Windows RCE in the 2019 list, while Microsoft lists $100,000 for unauthenticated non-sandboxed RCE in its Windows bounty.
Source ↗ - Geminiplace 5
Human bounties pay $100,000 to $250,000 for elevation of privilege, while agent clusters uncover deep win32k/NTFS race conditions for under $2,200.
- Kimiplace 6
Crowdfense lists a Windows zero-click full chain at up to $1M while Microsoft paid $20M+ in bounties in one year — and no public AI pipeline cost exists for the Windows kernel yet.
Source ↗ - DeepSeekplace 4
Hyper-V pays $250,000 for a working exploit; a Windows LPE sells for $150k-$300k on the gray market, while AI cost stays unpublished.
Source ↗
5 AIs put it here - 5
VMware ESXi and Hypervisors
Pwn2Own Berlin 2026 paid $150,000 for ESXi plus a $50,000 cross-tenant bonus. The pipeline also found a QEMU flaw in its own harness. No AI cost is published.
Included it: ChatGPT, Grok, Gemini, Kimi, DeepSeek. Did not include: Meta, Nemotron, Nova, MiMo, Wenxin, Qwen.In their lists: ChatGPT: 5 · Grok: see list · Gemini: 4 · Kimi: 7 · DeepSeek: 8
What each AI noted 5
- ChatGPTplace 5
The 2026 Berlin competition offered $150,000 for the virtualization target, with a conditional $50,000 cross-tenant bonus.
Source ↗ - Grok
Named differently in its own list: open the full list.
see list - Geminiplace 4
A single hypervisor guest-to-host escape commands $200,000 at Pwn2Own, while LLM-assisted symbolic solvers locate hypercall parsing faults for under $4,000.
- Kimiplace 7
The pipeline found a QEMU TCG vulnerability in its own test harness — traced with zero access to QEMU source — a hint that the virtualization layer under the cloud is in the blast radius.
- DeepSeekplace 8
VMware ESXi cross-tenant escape earned $200,000 at Pwn2Own — a hypervisor target where AI evidence is still thin.
Source ↗
5 AIs put it here - 6
Google Android
Google raised its top reward to $1.5M for hard Titan M2 chains while cutting rewards for flaws AI made easier to find. Zerodium lists up to $2.5M.
Included it: Meta, Gemini, Kimi, DeepSeek. Did not include: ChatGPT, Grok, Nemotron, Nova, MiMo, Wenxin, Qwen.In their lists: Meta: 5 · Gemini: 3 · Kimi: 3 · DeepSeek: 5
What each AI noted 4
- Metaplace 5
Google raised the top to $1.5M for a Titan M2 zero-click chain with persistence, still below Zerodium's $2.5M overall ceiling.
Source ↗ - Geminiplace 3
Zero-click full chains valued up to $5,000,000 on broker desks can be deconstructed by AI orchestrations analyzing open Android AOSP trees for roughly $6,500 in token passes.
- Kimiplace 3
The single most expensive software target on Earth — up to $2.5M at Zerodium and $5M at Crowdfense — has no public AI-cost baseline yet, which is precisely why the gap is scariest.
Source ↗ - DeepSeekplace 5
Android's top zero-click chain pays $1.5M, but Google cut some awards as AI made analysis cheaper.
Source ↗
4 AIs put it here - 7
WhatsApp and Messaging Apps
Pwn2Own offered $1M for a zero-click WhatsApp exploit, but the entry was withdrawn and Meta called the bugs low-risk. No AI cost is published.
Included it: Meta, Gemini, Kimi, DeepSeek. Did not include: ChatGPT, Grok, Nemotron, Nova, MiMo, Wenxin, Qwen.In their lists: Meta: 8 · Gemini: 7 · Kimi: 5 · DeepSeek: 7
What each AI noted 4
- Metaplace 8
Pwn2Own offered $1M for a zero-click WhatsApp RCE, the entry was withdrawn, and Meta called the disclosed bugs low-risk.
Source ↗ - Geminiplace 7
Brokers offer up to $5,000,000 for remote messenger compromise, while multi-agent parser auditing identifies zero-click heap corruptions for $5,000 in batch tokens.
- Kimiplace 5
Messaging zero-clicks pay $1M-$1.5M precisely because they are rare — the same scarcity logic that collapses if an agent pipeline ever reproduces one for five figures.
Source ↗ - DeepSeekplace 7
WhatsApp zero-click RCE+LPE is a $1.5M broker item; AI has no published price tag for it yet.
Source ↗
4 AIs put it here - 8
OpenSSL
An AI system found all 12 zero-days in OpenSSL's January 2026 release, including a CVSS 9.8 CMS flaw. Cost per bug isn't published.
Included it: Meta, Gemini. Did not include: ChatGPT, Grok, Nemotron, Nova, MiMo, Kimi, DeepSeek, Wenxin, Qwen.In their lists: Meta: 4 · Gemini: 10
What each AI noted 2
- Metaplace 4
AI found all 12 zero-days in OpenSSL's January 2026 release, including a CMS flaw rated 9.8.
Source ↗ - Geminiplace 10
Foundational C libraries that protect global finance can have subtle buffer overflows and state-machine flaws uncovered by LLM symbolic reasoning for less than $300.
2 AIs put it here - 9
DeFi Smart Contracts
Agents scanned a contract for $1.22 on average and built $4.6M in simulated exploits, against Immunefi caps of $5M.
Included it: Meta, Kimi. Did not include: ChatGPT, Grok, Nemotron, Gemini, Nova, MiMo, DeepSeek, Wenxin, Qwen.In their lists: Meta: 10 · Kimi: 10
What each AI noted 2
- Metaplace 10
Agents scanned a live contract for $1.22 on average and collectively built $4.6M in simulated exploits, in a market where Immunefi caps reach $5M.
Source ↗ - Kimiplace 10
The mean critical pays $114,355 against a $20,000 median — a heavy tail where one AI-found bug can be a seven-figure event, as the $10M Wormhole award proved.
Source ↗
2 AIs put it here - 10
FFmpeg
21 zero-days for about $1,000, roughly $48 each. The figure is vendor-reported by depthfirst and not independently reproduced.
Included it: Meta. Did not include: ChatGPT, Grok, Nemotron, Gemini, Nova, MiMo, Kimi, DeepSeek, Wenxin, Qwen.In their lists: Meta: 3
What each AI noted 1
- Metaplace 3
21 zero-days for about $1,000, roughly $48 per zero-day in about 1.5 million lines of C.
Source ↗
1 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
The $18.9M–$66.3M valuation of the 948 bugs applies historical contest prizes. It is hypothetical, not actual payouts, and not every finding would qualify.
kernelCTF pays one static $71,337 since 4 Sep 2026. Older '+$30,000 bonus' figures are retired. The 45x multiple is hypothetical: first-submission rules apply and OpenAI dropped SLAB_VIRTUAL from the test kernel.
Google repriced in opposite directions: standard Chrome payouts fell because AI made those bugs cheaper, while Android top rewards rose for the hardest Titan M2 chains.
Headline AI bug counts need auditing: Greg Kroah-Hartman counts only 10 real fixes among 79 Linux CVEs credited to Anthropic's Mythos. Validation, not generation, is the bottleneck.
03How we counted
In plain words, no formulas.
1The question
Top 10 Critical Software Targets Where AI Disrupted the Economics of Zero-Day Bounties
2Where it applies
global / cybersecurity & financial asymmetry
3What we looked at
- Inference and compute cost required by an AI agent pipeline to discover and validate a functional root/RCE exploit
- Commercial bug bounty and zero-day broker payout price for equivalent unpatched vulnerabilities
- OpenAI, 'A thousand fewer paths to root' — API cost, counts and per-exploit cost of the autonomous Linux kernel LPE pipeline, plus the kernelCTF run
- Google kernelCTF rules — Static reward ($71,337) and eligibility for hardened-kernel LPEs
- XBOW / Business Wire and CSA Labs — Compute cost (~$40,000) of an AI-built Chrome full chain that earned $250,000
- Apple Security Bounty — Published reward ceilings for exploit chains
- Microsoft MSRC Hyper-V bounty — Hyper-V guest-to-host reward tiers
- Trend Micro ZDI, Pwn2Own Berlin 2026 — Published contest prizes for virtualization, browsers and enterprise apps
4Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: ChatGPT, Meta, Grok, Nemotron, Gemini, Nova, MiMo, Kimi, DeepSeek, Wenxin, Qwen.
5How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
04Each AI's own list
🇺🇸 ChatGPT10 places · 4 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1Linux Kernel — 2026
948 distinct privilege-escalation vulnerabilities demonstrated in one $1.21 million research campaign.
Source ↗ - 2Microsoft Hyper-V — Windows Server and Azure Hosting
Microsoft publicly offers up to $250,000 for a qualifying guest-to-host compromise.
Source ↗ - 3Google Chrome — Browser and Sandbox
A $250,000 full-chain bonus coexists with much smaller rewards for individual memory-corruption findings.
Source ↗ - 4Apple iOS and macOS — Security Bounty
Apple's largest publicly advertised exploit-chain rewards exceed $5 million with bonuses.
Source ↗ - 5VMware ESXi — Enterprise Virtualization
The 2026 Berlin competition offered $150,000 for the virtualization target, with a conditional $50,000 cross-tenant bonus.
Source ↗ - 6Microsoft 365 — Word, Excel, PowerPoint and Outlook
A $150,000 application prize can be supplemented by a $50,000 Copilot-related award.
Source ↗ - 7Microsoft Edge — Chromium Browser
Pwn2Own Berlin valued a qualifying sandbox escape at $175,000 versus $75,000 for renderer-only compromise.
Source ↗ - 8Apple Safari — macOS Browser
The same 2026 competition priced renderer compromise at $75,000 and kernel escalation at $175,000.
Source ↗ - 9Linux KVM — Virtualization
A $50,000 competition target shows that the price of a virtualization escape is not the same as a Linux kernel LPE.
Source ↗ - 10Mozilla Firefox — Browser Security
A qualifying renderer compromise carried a $50,000 prize, doubling to $100,000 for specified escape or escalation outcomes.
Source ↗
What ChatGPT noticed
OpenAI reproduced 1,724 of 2,308 candidate Linux findings, approximately 74.7%. That validation rate matters more than raw scanner alerts.
The Linux pipeline identified 948 distinct LPE vulnerabilities but produced 1,169 working LPE exploits. Exploit count and unique-bug count are different economic denominators.
OpenAI's earlier July campaign found 166 distinct LPEs for $513,690. The October campaign found 948 for $1,213,430, showing a substantial improvement in measured research yield.
The strongest defensive implication is faster verification and patching, not that every remotely reachable system can now be compromised for approximately $1,000.
🇺🇸 Meta10 places · 4 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Linux Kernel (2026)
948 working root exploits from one $1.21M run, about $1,279 per unique LPE against $20,000 to $70,000 in historical contest value.
Source ↗ - 2Google Chrome (2026)
One researcher plus public GPT-5.6 Sol built a 3-bug full chain in 6 days for about $40,000 and took Google's first $250,000 full-chain reward.
- 3FFmpeg (2026)
21 zero-days for about $1,000, roughly $48 per zero-day in about 1.5 million lines of C.
Source ↗ - 4OpenSSL (2026)
AI found all 12 zero-days in OpenSSL's January 2026 release, including a CMS flaw rated 9.8.
Source ↗ - 5Android - Pixel Titan M2 (2026)
Google raised the top to $1.5M for a Titan M2 zero-click chain with persistence, still below Zerodium's $2.5M overall ceiling.
Source ↗ - 6Apple iOS - iPhone (2026)
Apple doubled its zero-click chain to $2M, with bonuses over $5M, matching broker money for the first time.
Source ↗ - 7Microsoft Windows (2026)
A broker paid $1M for zero-click Windows RCE in the 2019 list, while Microsoft lists $100,000 for unauthenticated non-sandboxed RCE in its Windows bounty.
Source ↗ - 8WhatsApp - Secure Messaging (2026)
Pwn2Own offered $1M for a zero-click WhatsApp RCE, the entry was withdrawn, and Meta called the disclosed bugs low-risk.
Source ↗ - 9AI Coding Agents and AI Infrastructure - Codex, LiteLLM (2026)
The AI tools themselves fell for $40,000 each at Pwn2Own Ireland 2026, Codex with a single argument-injection bug.
Source ↗ - 10Smart Contracts - DeFi (2026)
Agents scanned a live contract for $1.22 on average and collectively built $4.6M in simulated exploits, in a market where Immunefi caps reach $5M.
Source ↗
What Meta noticed
OpenAI says its reported costs are an upper bound, not a floor: sufficient harness engineering could meaningfully reduce cost at each stage.
Anthropic reported Claude Opus 4.6 found more than 500 validated zero-days in production open-source code, including a 23-year-old Linux kernel NFS bug.
Google moved Chrome and Android bounties in opposite directions in 2026: Chrome standard payouts fell as AI made bugs cheaper, Android top payouts rose for the hardest Titan M2 chains.
The FFmpeg run cost about $1,000 for 21 zero-days, which depthfirst says is 10 percent of what Anthropic spent using Mythos for a similar task.
🇺🇸 Grok10 places
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1Linux kernel (stock desktop distributions)
$1,279 AI cost vs $20k–$70k contest payouts yields 16–55× asymmetry on 948 distinct root paths
- 2Google kernelCTF hardened target
21 AI LPEs for ~$33k against $71,337 human bounty and only 26 human 0-days accepted all year
- 3Linux kernel (July 2026 GPT-5.6 Sol baseline)
Earlier model already delivered 166 unique LPEs at $3,094 each before the five-fold jump
- 4Linux kernel information-leak class
291 working leak PoCs produced alongside the root exploits in the same pipeline run
- 5Firefox (OpenAI cyber model)
81 previously unknown high-severity flaws surfaced in the same model family that produced the kernel haul
- 6QEMU TCG (harness discovery)
Pipeline found and exploited a QEMU TCG flaw in its own test harness without source access
- 7Linux kernel (novel exploitation techniques)
At least 10 techniques previously unknown to OpenAI researchers and external kernel experts
- 8Linux kernel (candidate-to-exploit funnel)
2,308 candidates → 1,724 reproduced → 1,169 working LPE exploits before variant collapse to 948
- 9Linux kernel (Patch the Planet June baseline)
GPT-5.5 Cyber earlier produced only 24 LPEs, illustrating the rapid capability jump
- 10Desktop Linux distributions with standard mitigations
All 948 exploits worked on stock desktops with KASLR, SMEP and SMAP enabled
🇺🇸 Gemini10 places
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1Linux Kernel (Master Branch / Core OS)
An autonomous agent pipeline extracted 948 root-level LPE exploits at $1,038 per weaponized bug against a target where human market bounties traditionally claim $70,000 to $100,000.
- 2Apple iOS (Zero-Click Mobile Chains & WebKit Escapes)
The $5,000,000 commercial zero-click bounty ceiling faces sub-$20,000 multi-agent reasoning chains that decompose sandbox escape primitives into automated AST fuzzing steps.
- 3Google Android (Baseband, OS Kernel & Pixel Hardened Stack)
Zero-click full chains valued up to $5,000,000 on broker desks can be deconstructed by AI orchestrations analyzing open Android AOSP trees for roughly $6,500 in token passes.
- 4VMware ESXi & Workstation (Virtualization Escapes)
A single hypervisor guest-to-host escape commands $200,000 at Pwn2Own, while LLM-assisted symbolic solvers locate hypercall parsing faults for under $4,000.
- 5Microsoft Windows (Host Kernel EoP & Hyper-V)
Human bounties pay $100,000 to $250,000 for elevation of privilege, while agent clusters uncover deep win32k/NTFS race conditions for under $2,200.
- 6Google Chrome (V8 Engine & Chromium Sandbox Escapes)
Broker payouts reach $3,500,000 for browser RCE and sandbox escape chains that autonomous agent pipelines now reconstruct from commit logs for roughly $3,000.
- 7WhatsApp & Apple iMessage (Zero-Click Messaging RCE)
Brokers offer up to $5,000,000 for remote messenger compromise, while multi-agent parser auditing identifies zero-click heap corruptions for $5,000 in batch tokens.
- 8Apple macOS & Safari (WebKit to Root Chains)
Apple's $1,000,000+ defense threshold meets AI-driven differential fuzzers that isolate macOS Mach portIPC race conditions for under $2,000.
- 9Enterprise Network Gateways (Ivanti, Fortinet, PAN-OS)
Edge appliances that yield $100,000 to $300,000 bounties for unauthenticated RCE can be systematically reverse-engineered by AI agent swarms for $450 in firmware analysis.
- 10OpenSSL & Core Cryptographic Libraries
Foundational C libraries that protect global finance can have subtle buffer overflows and state-machine flaws uncovered by LLM symbolic reasoning for less than $300.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1Linux kernel LPEs (2026)
One 30-hour, $1.2M run minted 948 root exploits at $1,279 each — prices that humans only get paid $20,000-$70,000 for at Pwn2Own and TyphoonPWN.
Source ↗ - 2Google Chrome exploit chains (2026)
A three-bug full Chrome chain that pays $250,000 was assembled by four people and AI agents in six days on ~$40,000 of compute — work that used to take a team 6-9 months.
- 3Android zero-click full chains (Zerodium)
The single most expensive software target on Earth — up to $2.5M at Zerodium and $5M at Crowdfense — has no public AI-cost baseline yet, which is precisely why the gap is scariest.
Source ↗ - 4iOS zero-click exploits (Apple)
Apple raised its own ceiling to $2M per chain (above $5M with bonuses) just as agents learned to chain sandbox escapes elsewhere — the premium target is one public AI-cost datapoint away from repricing.
Source ↗ - 5WhatsApp / iMessage RCE+LPE (Zerodium)
Messaging zero-clicks pay $1M-$1.5M precisely because they are rare — the same scarcity logic that collapses if an agent pipeline ever reproduces one for five figures.
Source ↗ - 6Windows exploit chains (Crowdfense / Microsoft)
Crowdfense lists a Windows zero-click full chain at up to $1M while Microsoft paid $20M+ in bounties in one year — and no public AI pipeline cost exists for the Windows kernel yet.
Source ↗ - 7QEMU / hypervisor stack (2026)
The pipeline found a QEMU TCG vulnerability in its own test harness — traced with zero access to QEMU source — a hint that the virtualization layer under the cloud is in the blast radius.
- 8curl / ubiquitous open-source C libraries (2026)
The first target whose economics inverted completely: AI slop made the bounty unaffordable, so curl now pays $0 — the bounty died of abundance.
Source ↗ - 9OWASP-class web applications (HackerOne, 2026)
An agent on a $5,000-a-month budget out-ranked every human on a HackerOne leaderboard — while the median real payout for this bug class is $500-$2,500.
Source ↗ - 10DeFi smart contracts (Immunefi, 2026)
The mean critical pays $114,355 against a $20,000 median — a heavy tail where one AI-found bug can be a seven-figure event, as the $10M Wormhole award proved.
Source ↗
What Kimi noticed
The unit economics flipped in public view: $1,279 of API credits per distinct Linux root exploit against human contest rates of $20,000-$70,000 — a 16x-55x asymmetry OpenAI itself prices at $18.9M-$66.3M for one 30-hour run.
One automated scan nearly matched an entire human year: 21 kernelCTF LPEs in 112 hours vs 26 confirmed human 0-day LPEs accepted by kernelCTF in all of 2026.
Hardening is losing the race: the October run found 5.7x more LPEs than the July run even though the kernel absorbed three months of additional hardening in between.
The defensive response is to cut prices, not raise them: curl killed its bounty ($0 now) and HackerOne's Internet Bug Bounty cut critical rewards from $9,250 to $2,257 — bounty floors are compressing toward AI discovery cost.
🇨🇳 DeepSeek10 places · 4 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1Linux kernel LPE (2026)
$1,279 buys a root exploit that kernelCTF pays $71,337 for — a 56x gap on the same bug class.
Source ↗ - 2Google Chrome full-chain exploit (2026)
$40,000 of AI compute earned a $250,000 Chrome full-chain bounty — the gap narrows when the target is a browser, not a kernel.
Source ↗ - 3Apple macOS (2026)
Apple's ceiling is $5M with bonuses, but AI found 50+ macOS flaws in three weeks — and Apple capped submissions.
Source ↗ - 4Microsoft Windows / Hyper-V (2026)
Hyper-V pays $250,000 for a working exploit; a Windows LPE sells for $150k-$300k on the gray market, while AI cost stays unpublished.
Source ↗ - 5Android (2026)
Android's top zero-click chain pays $1.5M, but Google cut some awards as AI made analysis cheaper.
Source ↗ - 6Apple iOS (2026)
iOS zero-click chains reach $5M at Apple and $9M at Crowdfense — the highest human price, with no published AI cost.
- 7WhatsApp zero-click RCE+LPE (2026)
WhatsApp zero-click RCE+LPE is a $1.5M broker item; AI has no published price tag for it yet.
Source ↗ - 8VMware ESXi (2026)
VMware ESXi cross-tenant escape earned $200,000 at Pwn2Own — a hypervisor target where AI evidence is still thin.
Source ↗ - 9Oracle Autonomous AI Database (2026)
Oracle's AI database had a $40,000 Pwn2Own target; organizers ran out of slots before every working RCE could be demonstrated.
- 10Microsoft Exchange RCE (2026)
Microsoft Exchange RCE earned $200,000 at Pwn2Own, but AI's cost to find it is not disclosed.
Source ↗
What DeepSeek noticed
OpenAI's own footnote says the $18.9M-$66.3M valuation 'does not represent actual bounty payouts.' Nobody paid anything; the markup is hypothetical.
The 948 is a patch count, not a CVE count: 1,169 working LPE exploits were grouped into 948 distinct patches, and kernel maintainers still review every fix.
kernelCTF's first-to-file rule means AI can find 948 bugs, but only the first submission per target is paid — so the aggregate bounty is capped by uniqueness, not volume.
Chrome's AI-assisted chain cost ~$40,000 and earned $250,000; the 6.25x gap is smaller than Linux's 56x, showing target complexity still protects some software.