Google finds automated AI patch-to-exploit framework | Magica
Google says hackers automated the path from security patch to exploit
Editorial Team
••📖7 min read
Google found an attacker framework configured to compare software patches, develop exploits and execute them without human intervention. The finding sharpens the pressure on defenders to patch quickly, but does not establish that entire attacks now run autonomously.
Google says it found an attacker framework that turns patch analysis into exploit development and execution without a human in that workflow.
The immediate risk centers on known vulnerabilities in systems that have yet to receive available fixes.
Earlier research and attacks already demonstrated substantial autonomy; the new claim is narrower than the arrival of wholly autonomous hacking.
Google’s threat researchers say attackers have connected a valuable sequence of tasks: reading software changes, identifying the weakness those changes repair, and building and running an exploit. That could accelerate the race between organizations installing fixes and attackers using the fixes as clues.
Google’s overview of multi-agent offensive frameworks, illustrating the broader architecture rather than an execution trace of exploit_agent. Source: Google Cloud.
A patch becomes an attacker’s research brief
The Google Threat Intelligence Group, Google’s threat-research team, said in its November 16 AI Threat Tracker report that it had, for the first time in its tracking, identified attackers using large language model frameworks to fully automate vulnerability discovery and exploitation without a human in that loop. The findings draw on threat tracking, defenses across Google’s platforms and frontline incident-response investigations by Mandiant, the cybersecurity business Google acquired in September 2022. Mandiant brought an established threat-intelligence and incident-response practice into Google Cloud.
The central example is exploit_agent, a multi-agent vulnerability-research and weaponization framework found on an exposed attacker staging server. It connects to multiple frontier and locally run language models and is configured to continuously compare versions of popular web applications around patch announcements. The pipeline analyzes the changes for vulnerabilities, develops exploits and executes them automatically, according to Google.
The server also held stolen victim data and scripts for internal reconnaissance, brute-forcing and installing webshells—programs used to retain remote access to compromised websites. Some scripts removed other attackers’ webshells. Google said victim telemetry suggested an interest in French media, entertainment and education websites. Russian-language configuration comments offer a clue about the tooling, rather than proof of the operator’s nationality.
Manufacturers, utilities and miners linked to data center construction have shed $4.2 trillion in market value since June. Strong supplier orders contrast with financing concerns, project delays and political resistance.
Editorial Team
The mechanism favors attacks against disclosed vulnerabilities, often called n-days. A patch supplies a bounded research problem: what changed, and what could the old version be made to do? Google argues that this can be more efficient than searching broadly for previously unknown, or zero-day, flaws. Organizations still running the old code are the exposed population.
This is a change in how attackers organize work as much as in the underlying attack. Google says the mechanics still include familiar weaknesses such as compromised credentials and unpatched systems. Automation can connect the stages and repeat them across targets while reducing the human labor needed between steps.
Magica chart of researcher-reported results from a selected 2024 sandbox benchmark. These percentages do not measure internet-wide attack success or exploit_agent’s performance. Source: Autonomous exploitation research.
The novelty has limits
Autonomous exploitation predates this finding. An April 2024 research paper tested an agent using OpenAI’s GPT-4 language model on 15 vulnerabilities reproduced in sandboxed systems. Given vulnerability descriptions, it exploited 13 of the 15 within five attempts, about 87%; success across individual runs was 40%. Removing the descriptions reduced the within-five-attempt result to about 7%. This selected, reproducible benchmark established technical precedent, rather than an internet-wide attack success rate.
The nine other language models tested achieved no successful exploits. Those were 2024 comparisons, not a ranking of today’s alternatives. The researchers calculated $8.80 in model API charges per successful exploit from a $3.52 average run cost and the 40% success rate, using prices at the time. That historical experiment does not establish the cost of operating Google’s newly discovered framework.
There is also a prior operational claim. In November 2025, AI developer Anthropic said attackers had manipulated its Claude Code coding tool in an espionage campaign detected that September. Anthropic attributed the operation with high confidence to a Chinese state-sponsored group. It said roughly 30 organizations were targeted, including technology companies, financial institutions, chemical manufacturers and government agencies, with success in a small number of cases.
Anthropic estimated that AI performed 80–90% of the campaign. Humans still selected targets, built the framework and intervened at critical decisions; Claude sometimes invented credentials or mislabeled public information as stolen secrets. The claim was extensive automation with occasional human direction, not the disappearance of operators. Google’s finding concerns a particular patch-analysis and exploitation loop.
Google’s report places exploit_agent within a broader adoption of open-source penetration-testing frameworks—tools intended to assess security that attackers can also use. It says it has observed CAIRN and PentAGI deployed in the wild, alongside experimentation with CyberStrikeAI and HexStrike AI. Google observed CAIRN in intrusion activity against targets including the U.S. aerospace sector, and a customized PentAGI platform conducting reconnaissance and vulnerability scanning across organizations in Asia between late August and early October 2026. These systems can connect models to security tools, generate payloads and adjust tactics using execution feedback. Their availability means attackers can assemble capabilities from existing software rather than build every component themselves.
But Google also describes a trade-off: unrestricted autonomous execution creates conspicuous traffic and recognizable code traces. Models can invent details or even targets. The report expects criminals willing to sacrifice stealth for reach to have stronger incentives to adopt such tools than espionage operators seeking prolonged, quiet access. That is Google’s assessment, not an established adoption rate.
Defenders face speed and verification costs
Generated code does not necessarily work. Google’s report, citing outside security research, describes hundreds of purported but nonfunctional proof-of-concept exploits, likely produced with LLM assistance, circulating after PHP patches released on October 21, 2026. Defenders had to verify or debunk them while looking for genuine exploitation. A demonstration video and early patched code had already appeared before the official fixes, giving both researchers and attackers material to analyze. The relevant defensive clock can therefore start before the formal patch release.
Automation also needs model access and compute. Google identifies access costs, token rate limits and geographic restrictions as hurdles to scaling. In separate investigations, Mandiant observed attackers stealing cloud credentials, consuming victims’ model quotas and provisioning powerful GPU instances to run their own tooling, including agentic frameworks used for further exploitation. This is evidence of infrastructure abuse, not a cost measurement for exploit_agent.
Google is also an interested participant. It develops the Gemini models, operates cloud infrastructure and sells security services. The same report promotes its automated defenses and describes account disruptions and model safeguards. Its incident visibility gives the findings weight; its defensive-product claims remain vendor claims rather than an independent comparison of protection or cost.
The report does not provide a measured runtime, operating cost or success rate for exploit_agent, or quantify how much it shortened the patch-to-exploitation interval. Stolen data on the staging server establishes a connection to harmful activity, but does not by itself measure how much of an entire intrusion was autonomous.
Google’s company-reported LLMJacking traffic illustrates separate cloud-resource abuse, not exploit_agent’s operating cost or performance. Source: Google Cloud.
What would establish the scale of the shift
The next evidentiary test is repeatability: how often these pipelines produce working exploits, how quickly they do so after a patch, and how many systems they successfully compromise. A report on Google’s findings attributes to its researchers a projection that sophisticated, fully autonomous attacks could emerge within six months. That is a forecast, not an observed result. Published measurements of success, elapsed time and human intervention would test it.
For defenders, the immediate decision is how quickly they can identify exposed systems and deploy available fixes. The evidence points to automation of the work attackers perform while those systems remain unpatched; it does not establish that human direction has disappeared from the wider operation.
RIKEN has stopped Fugaku’s research jobs after an intrusion into its operational network. Up to 4,597 users’ account information may be affected, and intruders tried acquired credentials on other systems; research-result leakage remains unconfirmed.
DeepSeek users report daily caps, shorter inputs and restrictions on resuming inactive conversations. Reported financing and a planned data center do not establish when more capacity will reach the free chat service, whose quota rules remain unclear.
Corma’s $125 million Series A brings total funding to $185 million as the defensive cybersecurity AI lab confronts a computing bottleneck. Its specialized-model strategy faces existing security agents and an unresolved test of performance and deployment costs.
Ask Maps is adding shareable travel schedules, saved-list availability searches and U.S. appointment finding. The update extends an existing AI planning service rather than introducing trip planning to Maps.
SemiAnalysis has reportedly cut its 2030 global AI accelerator spending forecast by about 40% to $923 billion, with Nvidia facing the largest revisions. Lower price assumptions and financing constraints help explain the reduction, while customer deals show both the strength and financial cost of competing chip platforms.
Federal Reserve Governor Michael S. Barr says banks report AI agents attempting account actions without verifiable permission. Autonomous deposit transfers could accelerate future bank runs, but their scale and speed remain unmeasured in his speech.
Amazon-owned Zoox is opening its Las Vegas robotaxis to Uber riders while keeping its own app. The November launch follows an earlier summer target; a joint MGM Resorts agreement addresses resort pickups, while fleet size and financial terms remain undisclosed.
Germany’s reported military AI shortlist excludes Palantir, while France and the Netherlands pursue European alternatives. The challenge is to secure national control over data and infrastructure while delivering usable systems that work with allies already adopting the US supplier.
China aims to make school AI education broadly available by 2030. National guidance already calls for pupil assessment, but an investigation finds common measures of AI literacy still lag an uneven rollout spanning courses, tutoring and student monitoring.
Drive Capital led Zus Health's $33 million Series B, extending a shared-record strategy backed by earlier financing and EHR partnerships. The company is expanding AI workflows and pursuing a more direct role in national data exchange, while disclosed activity figures leave customer returns unmeasured.
Atomarine’s $19 million seed round and Mocean Energy’s wave-powered computing proposal target land, water and electricity constraints. Earlier underwater projects show technical promise, but do not establish the costs or performance of these new floating designs.
Tencent’s new Hunyuan model charges 2 yuan per million output tokens and targets everyday office work. Its claimed 95% task-cost saving against Hy4 comes from an internal benchmark, with a trade-off on long coding tasks.
The Indian chip startup is funding programmable image sensors and targeting production in 2027. Its promised system savings remain unproven, while Sony's earlier designs challenge broad claims of novelty.
Baidu’s AI cloud infrastructure revenue nearly doubled in the third quarter, cushioning a 22% advertising decline. But AI applications were flat, app users fell, and the return to reported profit obscured lower adjusted earnings and a wider annual free-cash-flow deficit.
The Trump administration reportedly plans to designate more than two dozen AI companies to help win overseas business. Existing export rules support integrated technology packages, but guarantee neither financing nor licenses and leave room to support competing suppliers.
Singular leads the Danish accounting software company's seed round as it builds a London commercial hub and targets Benelux. Its pitch rests on automating preparation while accountants retain sign-off, with claimed efficiency gains still awaiting independent verification.
Alibaba says Qwen4-Max reduces average task costs by 62% on its workplace benchmark. The hosted launch builds on its predecessor’s agent capabilities as Alibaba expands a growing cloud business and bears the expense of AI infrastructure.
The FBI warns that ARTEX can accelerate intrusions, with some reaching privileged access in under three hours. Microsoft also documents unsuccessful mass targeting, while compromised backup-service credentials opened a separate route into customer environments.
Jensen Huang has reportedly discussed returning to Nvidia three days after resigning. Supporters want reinstatement during an investigation, while directors have installed interim leadership and begun a permanent CEO search.