AI agents are now turning patches into exploits, and the companies building AI models for security are gating their most capable versions

Google's Threat Intelligence Group reports that attackers have, for the first time, used AI agent frameworks to fully automate the process of comparing software patches, building working exploits and running them, with no human in the loop. In a separate case, Symantec describes an intrusion against a US retailer in which AI-generated PowerShell scripts achieved lateral movement within minutes after hand-typed commands had failed.

AI Security News 6 min read

AI agents are now turning patches into exploits, and the companies building AI models for security are gating their most capable versions

On 16 November 2026, Google's Threat Intelligence Group (GTIG) published its fourth-quarter AI threat tracker. The report says threat actors are using large language model (LLM) frameworks to compare code before and after a security update, identify the flaw the update fixes, write an exploit and run it automatically. GTIG says this is the first time it has observed fully automated vulnerability discovery and exploitation of this kind.

The report cites a suspected Russian-speaking operator who used an AI agent to turn patches into working exploits before stealing data from victim organisations. It also names several open-source frameworks, including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI, and describes financially motivated and suspected China-linked espionage actors experimenting with them.

The process has a name: differential patch analysis. An attacker compares the code before and after a fix, identifies what changed and builds an exploit for the unpatched systems that remain. GTIG also points to a campaign against PaperCut print management servers in August 2026, which GreyNoise says was built with LLM assistance and targeted hundreds of organisations.

The practical consequence is the one Google highlights. Once a patch is published, the time available to install it before exploitation begins may now be much shorter than the time many organisations take to test and deploy updates.

The US retailer intrusion

A separate case, described by Symantec on 12 November 2026, involved an unnamed US retailer targeted in September. According to Symantec, the attacker first used hand-typed commands, which were repeatedly blocked. When the attacker resumed on 22 September, the commands changed style. Symantec describes long, error-free PowerShell scripts with labelled output sections, and in places comments that resemble a model's reasoning.

About ten minutes after querying the retailer's SolarWinds Orion monitoring software, the attacker used a component bundled with it to run commands on a second server. The attacker harvested credentials and achieved lateral movement within minutes.

Symantec's conclusion is cautious. The timing between commands was more consistent with a human operator relaying output to an AI tool than with a fully autonomous agent. Symantec's assessment that AI was used is an inference from the scripts and timing, not something the attacker confirmed. Later in the same intrusion, five commands in a row failed with simple errors, which suggests the attacker still needed human judgement.

The new security models

On 5 November 2026, OpenAI released GPT-6 Cyber, a cybersecurity-specialised version of its GPT-6 model. OpenAI rates it "Critical" under its own Preparedness Framework, one level above the rating for its previous cyber model. Access is limited to vetted customers in its Daybreak Red programme, which provides offensive red-teaming access to advanced models. OpenAI reports that the model found 81 previously unknown vulnerabilities in Firefox, developed exploits for 948 Linux kernel vulnerabilities and built a working remote code execution exploit for Brotli, an open-source compression library. All reported findings were disclosed to maintainers and patched, according to OpenAI.

OpenAI describes the safeguards as identity and account security requirements, more frequent checks with OpenAI for high-risk requests, and monitoring for rule-breaking. In its own tests, the model attacked an off-limits test service in 0.07% of graded runs without its deployment safety prompt, and in none of the runs with it.

On 9 November 2026, Anthropic released Claude Mythos 5.2 to selected partners, alongside Claude Fable 5.2, which is available to all API customers. Anthropic says Mythos 5.2 scores 89.6% on its CVE Rediscovery benchmark, compared with 78.3% for Mythos 5.1, and that it is better at finding, reproducing and proposing fixes for vulnerabilities. Mozilla reports that the model roughly doubled its vulnerability discovery numbers in Firefox scans compared with Mythos 5.1. Anthropic notes that some of these gains come from a single tester's raw counts, so results will vary by codebase. Mythos 5.2 is also available through cloud platforms and some security products for approved customers.

Both companies' figures come from their own evaluations and from partners' reports. Neither has been independently replicated in full.

Why the two stories belong together

The Google report describes what attackers are already doing with AI. The retailer case shows what that looks like inside a network. The model releases show what the AI developers are doing about it: putting the most capable security models behind vetting, monitoring and identity checks, rather than releasing them publicly.

Gating addresses one risk, the direct use of a model as an attack tool. It does not address the other risk Google describes, in which open-source frameworks and general-purpose models are combined by attackers without any vendor's involvement. A restricted model keeps a particular capability out of reach of most users, but it does not prevent the same techniques from appearing in other tools.

What organisations should do

  1. Shorten patch deployment for critical and internet-facing systems. If attackers can build an exploit from a patch within hours, a monthly patch cycle is a long window. Prioritise by exposure, and define a target time for critical updates.

  2. Monitor for the behaviour, not just the tool. The retailer case involved ordinary PowerShell and a legitimate monitoring product. Look for unusual credential harvesting, unexpected script execution and rapid movement between servers.

  3. Watch SolarWinds Orion and other management platforms closely. Monitoring and management software often has broad access, and the retailer case used such a component to reach a second server.

  4. Assume scripts can be well written. Typos are no longer a reliable sign of manual work, so detection should rely on behaviour and context.

  5. Use defensive AI where it is available to you. If your organisation qualifies for access to restricted security models through a vendor programme, the potential for finding flaws before attackers do is significant. Check whether your vendors offer such access, and document how any model output is validated before it changes production systems.

The lesson

AI is shortening the gap between the publication of a vulnerability fix and its use in an attack. Google's report shows that this is no longer a theoretical concern. The retailer case shows that the speed is already reaching real intrusions. The vendors' restricted releases are a serious attempt to keep the most capable tools in the hands of defenders, but they cannot control every tool. Organisations should plan for attackers who move at machine speed, and should treat the patch window as the most urgent number in their security programme.

Blog

Security 4 min read

Four years after the Rockstar breach, the hacker pleads guilty: what the case shows about insider-style access and extortion

On 16 November 2026, Arion Kurtaj, 21, pleaded guilty at Woolwich Crown Court to hacking Rockstar Games in 2022, along with other attacks linked to the Lapsus$ group. Kurtaj reportedly gained access to Rockstar's internal Slack channels, stole around 90 clips of unreleased Grand Theft Auto VI footage, and attempted to extort the company. When Rockstar refused to pay, he published the material. He will be sentenced on 23 November 2026. Rockstar says recovery cost about $5 million.

Security 4 min read

Telegram widgets under scrutiny: Kaspersky warns of suspected WebRAT delivery

Kaspersky has warned that attackers may have used Telegram's newly launched widgets to deliver WebRAT malware. The warning followed public remote code execution (RCE) demonstrations by the security firm V12. Telegram CEO Pavel Durov has rejected the claims, saying no one could be hacked just by receiving a message and that the issue was fixed on Telegram's servers.

Security 4 min read

N-central under attack: 4EverAble2Shell exploit chain is installing crypto miners on unpatched and patched servers

Attackers are exploiting 4EverAble2Shell, an unauthenticated exploit chain against N-able N-central, to gain root access on on-premises servers and install cryptocurrency miners. Researchers reported at least 21 compromised servers within hours of a public proof-of-concept appearing. Nineteen of them were running the latest hotfix, so the fix alone is not a complete answer. If your organisation runs on-premises N-central with the Legacy API enabled, treat any exposed server as a priority.