← Voltar ao feed

GTIG: Actors adopt agentic AI frameworks; exploit_agent automates patch-to-exploit on French targets, SOMBERMEME via Claude Code

9 fonte(s) · 8 atualização(ões) · 1 repetição(ões) · publicado 16/11/2026 11:00 · coletado há 9 horas

Matéria consolidada. Reúne 9 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 9 fontes trouxeram em 17/11/2026 12:10; os links antigos trazem para cá.

Fontes: infosecurity-magazine.com ↗ helpnetsecurity.com ↗ gbhackers.com ↗ cyberscoop.com ↗ securityaffairs.com ↗ hackread.com ↗ thehackernews.com ↗ theregister.com ↗ bleepingcomputer.com ↗

Google Threat Intelligence Group (GTIG) and Mandiant report in the Q4 2026 AI threat tracker (published 2026-11-16) that financially motivated and state-sponsored actors are moving from manual chatbot prompting to autonomous multi-agent offensive frameworks that automate vulnerability discovery and exploitation. GTIG calls this a "paradigm shift" in the past quarter (infosecurity-magazine.com). It describes what it calls the first observed case of fully automated vulnerability research and exploitation with no human in the loop: exploit_agent, found on an exposed staging server next to stolen victim data (bleepingcomputer.com, theregister.com, hackread.com). Victim data points to French media, entertainment and education websites. Google did not name a group or state; Russian-language comments and configuration notes in the tooling led The Register and CyberScoop to describe the operator as suspected Russian-speaking (theregister.com, cyberscoop.com), while Hackread says those comments do not establish nationality (hackread.com). Open-source frameworks including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI are also being abused or tested. GTIG projects that unsophisticated, near-autonomous attacks will appear within three months and sophisticated, fully autonomous attacks within six (bleepingcomputer.com); Infosecurity Magazine reports the expectation as sophisticated, fully autonomous attacks within three to six months (infosecurity-magazine.com). Google disabled the associated accounts (hackread.com). None of the sources consulted the primary GTIG report directly (Infosecurity Magazine also worked from the article alone); The Hacker News worked from a report shared with it, so details are unverified against the original.

Entenda a falha

A vulnerabilidade

The story lists five CVEs, but the articles describe few of them technically. CVE-2026-40217 is a LiteLLM sandbox escape that escalates a limited key to admin; in observed cases the attacker held a limited key from stolen credentials. CVE-2026-81578 and CVE-2026-82078 are two PaperCut NG/MF flaws that GBHackers calls zero-days. Security Affairs says they were exploited as zero-days in late August, while The Hacker News places the PaperCut case in September. CVE-2026-109618 and CVE-2026-109619 are PHP flaws patched on October 21. The articles give no flaw type, affected versions or fixed versions for any of the five.

Como funciona

For CVE-2026-40217, the actor starts with a limited LiteLLM key taken from public GitHub repositories or fake session-token harvesting sites. LiteLLM should keep that key inside its restricted permissions within a sandbox. The actor escapes the sandbox and the key gains admin rights over the deployment. For the PaperCut and PHP flaws, the articles describe how they were weaponized, not how they work. LLM-driven pipelines compare code before and after a patch (including pre-release channels or early patched code), read release notes and PoC code, identify the fixed flaw and draft an exploit. The internal mechanism of the PaperCut and PHP bugs is not described.

Por que é vulnerável

The articles do not state the root cause of any of the five CVEs. For CVE-2026-40217 they describe only the effect: the sandbox does not hold, so a limited key becomes admin. They do not say what any fix changes. The broader point of the GTIG report is that published patches and diffs reveal the fixed flaw, so internet-exposed systems with low-complexity flaws stay exposed until updated.

Como foi descoberta

The articles do not say who first found any of the five flaws. CVE-2026-40217 surfaced through observed exploitation: Mandiant handled LLMjacking incidents in which an actor used it. The PaperCut flaws were exploited as zero-days. Per Security Affairs, Huntress and other researchers built working exploits within 24 hours and GreyNoise detected a campaign four days after disclosure; GBHackers says that campaign likely used LLMs. For the PHP flaws, Security Affairs says researchers used LLMs to analyse them and that patched-code details and a PoC video appeared before the October 21 patch; GBHackers also says early patched code let attackers compare versions.

PoC pública

For the PaperCut flaws, Security Affairs says Huntress and other researchers developed working exploits within 24 hours, and GBHackers says researchers produced them within a day; the articles do not say where they were published. For the PHP flaws, a PoC video and patched-code details appeared before the October 21 patch, followed by hundreds of purported exploits, mostly nonfunctional and likely LLM-generated, which forced defenders to check which were real. BleepingComputer ties AI-generated PoCs to PHP2Shell, but the sources do not confirm it is the same case. No article mentions a PoC for CVE-2026-40217. Given how fast exploits appeared, patching PaperCut and PHP is urgent.

Explicação gerada por IA a partir da matéria e dos artigos, em 17/11/2026 12:26.

Exploit público

Há código de exploração publicado no Exploit-DB para CVE-2026-109618. O código não é reproduzido aqui: o link leva à fonte, e a explicação foi escrita pela IA a partir do arquivo publicado.

  1. exploit-db.com · EDB-52704 ↗ · remoto · php · pub. 22/10/2026 verificado

    PHP2Shell (CVE-2026-109618) - PHP core versions after v5.4 - RCE

    This script attacks PHP running under Apache mod_php (prefork) by sending crafted multipart uploads to any PHP page, abusing session upload-progress handling to corrupt memory. It needs only network access and no login, but the target must have specific settings: upload progress enabled, non-strict sessions, a low memory limit and a single worker. It leaks memory addresses via session cookies, then makes PHP run a short system command (14 bytes max). It looks like a working but fragile exploit, tuned to PHP 8.4.26 builds; the claim of all versions after 5.4 is unproven.

    CVE-2026-109618 · autor: OffSec · explicação gerada por IA

Detalhes técnicos

  • exploit_agent chains multiple agents using several frontier and locally hosted LLMs. It continuously diffed popular web application code before and after security updates, identified the fixed flaw, drafted exploits and executed them automatically (theregister.com, bleepingcomputer.com, cyberscoop.com). No success rate is given, and it is not established that every generated exploit worked (hackread.com).
  • The staging server also held data stolen from victims, reconnaissance data, webshells and brute-forcing scripts. Post-compromise scripts covered internal reconnaissance, password guessing, webshell installation and removal of rival attackers' webshells. The tooling used Russian-language content, including a configuration file telling the user to enter API keys for "auto-exploitation" (theregister.com, bleepingcomputer.com, cyberscoop.com).
  • ARTEX manages a set of sub-agents that plan, execute and adapt across kill chain stages: reconnaissance and scanning, initial access, persistence and privilege escalation, internal reconnaissance and lateral movement, and completion of objectives (infosecurity-magazine.com). The open-source multi-agent frameworks CAIRN, CyberStrikeAI and PentAGI are growing in popularity, and some groups are building custom harnesses (infosecurity-magazine.com).
  • The frameworks share three core capabilities: finding and exploiting low-complexity vulnerabilities, using LLMs to generate single-use payloads on demand, and using real-time execution feedback to adjust tactics (infosecurity-magazine.com).
  • CAIRN was used in attacks on the U.S. aerospace sector (bleepingcomputer.com, securityaffairs.com). PentAGI was seen alongside proxy services against the aerospace sector (theregister.com) and was deployed in the wild (hackread.com). Between late August and early October, one actor used a modified PentAGI to scan organizations across Asia (securityaffairs.com). HexStrike AI and CyberStrikeAI were still in testing (securityaffairs.com, hackread.com).
  • Mandiant-documented ARTEX intrusions show agents running in bursts of roughly 2 to 25 seconds, covering Active Directory reconnaissance, privilege escalation and webshell deployment (gbhackers.com).
  • GTIG saw a PRC-nexus group building its own automated attack framework, likely for espionage (securityaffairs.com, cyberscoop.com). Financially motivated actors and suspected China-aligned espionage actors are experimenting with these frameworks (theregister.com, thehackernews.com).
  • Other tools named: ZeroHollow (inference service for exploit development and LLM routing) and Agent Zero (open-source agent framework with autonomous code execution) (thehackernews.com, securityaffairs.com).
  • LLMs compare source code versions, including pre-release channels, and read release notes and PoC code to weaponize patched n-day flaws. GTIG cites the PaperCut NG/MF flaws as an observed case (thehackernews.com). Per Security Affairs, two PaperCut flaws were exploited as zero-days in late August, working exploits appeared within 24 hours, and in one case attackers went from initial access to domain administrator within minutes (securityaffairs.com). GBHackers names the PaperCut zero-days CVE-2026-81578 and CVE-2026-82078, says researchers produced working exploits within a day, and says GreyNoise reported a campaign four days later that likely used LLMs to build exploits and reach domain admin within minutes (gbhackers.com).
  • Disclosures and exploits nearly doubled between January and August 2026. Zero-days averaged 11 per month in 2026 against 8 in 2025, and reached 22 in August (securityaffairs.com, cyberscoop.com, gbhackers.com).
  • PHP flaws CVE-2026-109618 and CVE-2026-109619 were patched on October 21. Many purported exploits were fake or nonfunctional and likely AI-generated, which burdened defenders (securityaffairs.com, cyberscoop.com). GBHackers adds that early patched code let attackers compare versions before the official patches, followed by hundreds of non-functional LLM-generated PoCs (gbhackers.com). BleepingComputer ties AI-generated PoCs to the PHP2Shell disclosure; the sources do not state that this is the same case.
  • LLMjacking: stolen LLM credentials and cloud resources are resold through proxy services and "LLM mixers" such as APIMart and JetNight, at $0.10 to $25 per million tokens. Mandiant saw fraudulent charges above $100,000 per day, with the worst case reaching $1 million. Actors used credentials from public GitHub repositories and fake session-token harvesting sites. One actor used CVE-2026-40217, a LiteLLM sandbox escape that escalates limited keys to admin (securityaffairs.com, thehackernews.com). ARTEX operators fed stolen LLM API keys back into their own infrastructure, creating a self-funding loop (gbhackers.com).
  • UNC6836, with suspected links to "Midnight Neptune", targeted cryptocurrency firms in North America and Europe. After initial compromise, it changed the settings of a modified Claude Code so that launching it downloaded the SOMBERMEME backdoor (bleepingcomputer.com). Security Affairs attributes fake recruiter profiles, ClickFix lures and edits to .claude/settings.json to MIDNIGHT NEPTUNE with the same result (securityaffairs.com). Hackread describes North Korea-aligned activity with AI personas targeting crypto businesses and attempts to manipulate coding assistants' configuration, without naming the group (hackread.com).
  • Other actors named: CALANQUE ION (formerly APT42) and LEAKY CASTLE (securityaffairs.com). The Chinese-language PhaaS platform Jieyou added an LLM chatbot that lets low-skilled affiliates customize phishing panels and lures (thehackernews.com, securityaffairs.com).
  • Other findings: AI-assisted influence operations, including a Gemini API-based platform meant to manage more than 100,000 accounts, and China-aligned actors using LLMs to process stolen data (hackread.com).
  • An FBI advisory warned that multiple actors have weaponized ARTEX since August against the financial, government, technology, defense and healthcare sectors, among others (bleepingcomputer.com; the advisory itself was not checked). Infosecurity Magazine says the FBI warned the previous week of growing ARTEX use, with South Korean financial institutions among the victims (infosecurity-magazine.com).
  • The report draws on Mandiant incident response telemetry, threat actor tracking and live platform defenses (infosecurity-magazine.com).
  • Targets are internet-exposed endpoints with low-complexity flaws that LLMs can easily exploit. The techniques are not new; the intrusion mechanics are still compromised credentials and unpatched systems, but the frameworks run them much faster, trading stealth for velocity (bleepingcomputer.com, securityaffairs.com, gbhackers.com, infosecurity-magazine.com). Because of this, attacks currently leave forensic signals and code-level traces, but GTIG expects stealth to improve (infosecurity-magazine.com).

Remediação

  • Google SecOps released detection rules for ARTEX, described as behavioral and volumetric rulepacks, and Mandiant published a hardening guide and defensive framework (securityaffairs.com, infosecurity-magazine.com).
  • Patch internet-exposed systems and web applications quickly, since the window between patch release and working exploit is shrinking (securityaffairs.com, hackread.com, cyberscoop.com). Patch known vulnerabilities and protect credentials (infosecurity-magazine.com).
  • Restrict access to applications that cannot be patched yet, and hunt for prior compromise, because patching does not evict an existing intruder (hackread.com).
  • Protect LLM and cloud credentials, and favor behavioral detection over single-string signatures, since many ARTEX indicators are LLM-generated (securityaffairs.com).
  • Exabeam's Sally Vincent recommends behavioral monitoring that flags unfamiliar tool use, unusual data access and changes in activity volume (infosecurity-magazine.com).
  • Mandiant recommends isolating control planes, replacing long-lived secrets with short-lived credentials, enforcing least privilege, segmenting networks and detecting behavioral anomalies, because static indicators miss agents that use valid credentials and legitimate features (gbhackers.com).

Relatos divergentes

  • Attribution of exploit_agent: The Register (theregister.com) and CyberScoop (cyberscoop.com) describe the operator as Russian-speaking, based on Russian-language comments and configuration in the tooling; CyberScoop says Google named no group or state. Hackread (hackread.com) says the Russian-language comments do not establish the operator's nationality. BleepingComputer (bleepingcomputer.com) does not attribute exploit_agent. No attribution is confirmed.
  • Midnight Neptune link: BleepingComputer links UNC6836 to "Midnight Neptune" only as a suspected link; Security Affairs (securityaffairs.com) attributes the Claude Code settings edits directly to MIDNIGHT NEPTUNE. Hackread mentions North Korea-aligned activity without naming the group. Neither source is primary.
  • Aerospace and Asia tooling: BleepingComputer and Security Affairs name CAIRN for the U.S. aerospace attacks. The Register names PentAGI, seen with proxy services, against the aerospace sector. Security Affairs places PentAGI in scanning across Asia, and CyberScoop lists CAIRN and PentAGI together against a U.S. aerospace organization and organizations across Asia. The assignment of each tool to each campaign is unresolved.
  • Framework list: PentAGI appears in The Register, The Hacker News, Hackread, Security Affairs, CyberScoop, GBHackers and Infosecurity Magazine, but not in BleepingComputer. HexStrike AI is described as abused in the earlier record but as still in testing by Security Affairs (the more recent and detailed source); GBHackers also describes only experimentation. Infosecurity Magazine lists ARTEX, CAIRN, CyberStrikeAI and PentAGI and does not mention HexStrike AI.
  • Autonomy forecast: BleepingComputer gives separate horizons (unsophisticated near-autonomous attacks within three months, sophisticated fully autonomous attacks within six); Infosecurity Magazine gives a combined three-to-six-month window for sophisticated, fully autonomous attacks. Neither is primary; the primary GTIG wording was not checked.
  • FBI advisory scope: BleepingComputer lists financial, government, technology, defense and healthcare sectors; Infosecurity Magazine mentions South Korean financial institutions among the victims and says the warning came the week before its article. The advisory itself was not checked, and the two descriptions are not necessarily incompatible.
  • PaperCut timing: The Hacker News says the PaperCut NG/MF flaws were exploited in September; Security Affairs says two flaws were exploited as zero-days in late August.
  • PHP flaw PoC timing: Security Affairs says PoC details appeared before the October 21 patch, and GBHackers says early patched code allowed version comparison before the official patches; CyberScoop says hundreds of purported exploits appeared within days after the patch.
  • Report date: The known report gives 2026-11-16 as the tracker's publication date, and Infosecurity Magazine also says November 16; GBHackers says the tracker is dated November 17.

Linha do tempo

  1. infosecurity-magazine.com ↗ · pub. 17/11/2026 06:30 · coletado 17/11 07:02

    Google GTIG warns fully autonomous agentic attacks are 3–6 months away as threat actors adopt ARTEX, CAIRN, CyberStrikeAI and PentAGI

    ler a análise desta fonte

    Google Threat Intelligence Group (GTIG) says in its quarterly GTIG AI Threat Tracker, published on November 16, that it saw a "paradigm shift" in the past quarter. Threat actors are using pen-testing frameworks such as ARTEX in intrusions. Google expects sophisticated, fully autonomous agentic attacks to emerge within three to six months. The report draws on Mandiant incident response telemetry, threat actor tracking and live platform defenses. This summary is based on the article alone: no web research was done, and the primary GTIG report was not opened.

    Technical Details

    • ARTEX manages a set of sub-agents. They plan, execute and adapt across kill chain stages: reconnaissance and scanning, initial access, persistence and privilege escalation, internal reconnaissance and lateral movement, and completion of objectives.
    • Open source multi-agent frameworks CAIRN, CyberStrikeAI and PentAGI are also growing in popularity. Some groups are building custom harnesses.
    • The frameworks share three core capabilities: they find and exploit low-complexity vulnerabilities, they use LLMs to generate single-use payloads on demand, and they use real-time execution feedback to adjust tactics.
    • The intrusion mechanics are unchanged: compromised credentials and unpatched systems. The speed and scale are what is new.
    • Attackers currently trade stealth for velocity, so attacks leave forensic signals and code-level traces. GTIG expects stealth to improve.
    • The article reports that the FBI warned the previous week of growing ARTEX use, with South Korean financial institutions among the victims.

    Remediation

    • Google SecOps has deployed behavioral and volumetric rulepacks to detect ARTEX.
    • Mandiant released a hardening guide and defensive framework.
    • Exabeam's Sally Vincent recommends behavioral monitoring that flags unfamiliar tool use, unusual data access and changes in activity volume.
    • Patch known vulnerabilities and protect credentials.
  2. gbhackers.com ↗ · pub. 17/11/2026 02:36 · coletado 17/11 04:00

    Google GTIG: Threat actors run autonomous multi-agent exploit pipelines (exploit_agent, CAIRN, PentAGI) with no human in the loop

    This article adds the CVE IDs for the PaperCut zero-days (CVE-2026-81578, CVE-2026-82078), a GreyNoise campaign four days after disclosure, ARTEX agent burst timing of about 2 to 25 seconds, and the detail that early patched PHP code let attackers diff versions before the official patches. It dates the report to November 17 and says the ARTEX frameworks were tied to a LiteLLM privilege escalation.

    ler a análise desta fonte

    This article adds the CVE IDs for the PaperCut zero-days (CVE-2026-81578, CVE-2026-82078), a GreyNoise campaign four days after disclosure, ARTEX agent burst timing of about 2 to 25 seconds, and the detail that early patched PHP code let attackers diff versions before the official patches. It dates the report to November 17 and says the ARTEX frameworks were tied to a LiteLLM privilege escalation.

    Google Threat Intelligence Group (GTIG) warns in its November 17 AI Threat Tracker that threat actors have moved from basic prompting to autonomous multi-agent frameworks that discover and exploit vulnerabilities without human intervention. GTIG found an exposed staging server hosting exploit_agent, which connects to frontier and local LLMs, compares code before and after patches, drafts exploits and runs them. Victim data suggested the operator targeted French media, entertainment and education websites, and Russian-language comments told the operator to enter LLM API keys for "auto-exploitation". The article also covers CAIRN (US aerospace), a customized PentAGI platform (Asia, late August to early October 2026), ARTEX, and LLMjacking monetization. These frameworks trade stealth for speed. The article is a secondary news report, and the primary GTIG report was not opened.

    Technical Details

    • exploit_agent also held scripts for internal reconnaissance, brute-forcing, webshell installation and removal of rival attackers' webshells.
    • ARTEX intrusions documented by Mandiant show agents running in bursts of roughly 2 to 25 seconds, covering Active Directory reconnaissance, privilege escalation and webshell deployment.
    • After the PaperCut zero-days CVE-2026-81578 and CVE-2026-82078, working exploits appeared within a day. GreyNoise reported a campaign four days later that likely used LLMs and reached domain admin within minutes.
    • Zero-day exploits reached 22 in August, against an average of 11 per month in 2026.
    • PHP flaws CVE-2026-109618 and CVE-2026-109619 saw early patched code diffed before the official patches, followed by hundreds of non-functional LLM-generated PoCs.
    • LLMjacking: fraudulent charges above $100,000 in a day and $1 million cumulatively. Attackers abused exposed cloud tokens and CVE-2026-40217 in LiteLLM, and resold access via "LLM Mixers" at $0.10 to $25 per million tokens. ARTEX operators fed stolen API keys back into their own infrastructure.
    • State-linked actors from China, Iran, Russia and North Korea use LLMs for reconnaissance, data analysis and influence operations.

    Remediation

    • Mandiant recommends isolating control planes, replacing long-lived secrets with short-lived credentials, enforcing least privilege, segmenting networks and detecting behavioral anomalies. Static indicators will miss much of this activity because agents use valid credentials and legitimate features.
  3. cyberscoop.com ↗ · pub. 16/11/2026 19:51 · coletado 16/11 21:00

    Google GTIG: Russian-speaking actors ran first fully automated AI attack chain ('exploit_agent') against French websites

    ler a análise desta fonte

    Google Threat Intelligence Group and Mandiant reported on Monday a quarterly finding that Russian-speaking actors ran what they assess to be the first intrusion operated end-to-end by AI, with no human in the loop. Google did not name a group or state. Attribution rests on Russian-language comments in the code, including a configuration file telling the user to enter API keys for "auto-exploitation". Victim data on the staging server pointed to French websites in the media, entertainment and education sectors. This summary rests on the CyberScoop article alone. No web research was done, so the Google report, CVE IDs and severity scores were not checked.

    Technical Details

    • The system, dubbed exploit_agent, chains multiple AI agents using frontier and locally hosted language models. It diffs widely used software before and after vendor patches, infers what the fix reveals, drafts an exploit and runs it automatically.
    • Investigators also found scripts for internal reconnaissance, password guessing and webshell installation. One script removed the webshells of rival attackers.
    • Google also saw open-source offensive frameworks CAIRN and PentAGI used against a US aerospace organization and organizations across Asia. It cited a China-nexus group building an automated attack framework, likely for espionage.
    • After two PHP flaws were patched on Oct. 21, hundreds of purported exploits appeared within days. Most were nonfunctional and likely AI-generated.
    • Google counted 22 exploited zero-days in August, against 11 per month earlier in 2026 and an average of 8 in 2025.

    Remediation

    The article gives no remediation. The implication for defenders is that the window between patch release and exploitation is shrinking, so patching should be prioritized accordingly.

  4. securityaffairs.com ↗ · pub. 16/11/2026 16:42 · coletado 16/11 17:00

    Google GTIG: attackers deploy autonomous AI agent frameworks (ARTEX, CAIRN, PentAGI) to automate n-day exploits and LLMjacking

    ler a análise desta fonte

    Google Threat Intelligence Group (GTIG) reports that threat actors have moved from manual chatbot prompting to autonomous multi-agent offensive frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI. GTIG observed CAIRN and PentAGI in real attacks, including CAIRN use against US aerospace organizations, and saw a PRC-nexus group building its own framework. Techniques are not new; the frameworks still depend on stolen credentials and unpatched systems, but they run them at far greater speed and trade stealth for velocity. The report also covers LLMjacking and "LLM mixers" that resell stolen LLM access, and the use of LLMs in information operations. This record rests on the article alone, because no web research was run.

    Technical Details

    • Frameworks automate patch diffing and n-day weaponization against internet-facing servers. GTIG found an exposed staging server running exploit_agent, which compares patched and unpatched web software, drafts exploits and runs them with no human in the loop. Victim data pointed to French media, entertainment and education sites, and the configuration notes were in Russian.
    • Between late August and early October, one actor used a modified PentAGI to scan organizations across Asia. HexStrike AI and CyberStrikeAI were still in testing. ZeroHollow (exploit development and LLM routing) and Agent Zero were also observed.
    • Disclosures and exploits nearly doubled between January and August 2026. Zero-days averaged 11 per month in 2026, against 8 in 2025, and reached 22 in August.
    • PaperCut: two flaws were exploited as zero-days in late August, and working exploits appeared within 24 hours. In one case, attackers went from initial access to domain administrator within minutes.
    • PHP flaws CVE-2026-109618 and CVE-2026-109619: PoC details appeared before the October 21 patch, and many exploit claims were fake.
    • LLMjacking: Mandiant saw fraudulent charges above $100,000 per day, with the worst case reaching $1 million. Actors used credentials from public GitHub repositories and fake session-token harvesting sites. One actor used CVE-2026-40217, a LiteLLM sandbox escape that escalates limited keys to admin. Stolen access is sold through LLM mixers at $0.10 to $25 per million tokens (APIMart, JetNight).
    • MIDNIGHT NEPTUNE used fake recruiter profiles, ClickFix lures and edits to .claude/settings.json so that starting Claude Code downloaded the SOMBERMEME backdoor. Other actors named: CALANQUE ION (formerly APT42), LEAKY CASTLE. The Jieyou phishing service added an LLM chatbot.

    Remediation

    Google SecOps released detection rules for ARTEX, and Mandiant published a hardening guide. Patch quickly, protect LLM and cloud credentials, and favor behavioral detection over single-string signatures, since many ARTEX indicators are LLM-generated.

  5. hackread.com ↗ · pub. 16/11/2026 14:24 · coletado 16/11 16:01

    Google GTIG: Threat Actor Uses 'exploit_agent' AI Framework to Turn Patches Into Exploits Autonomously

    ler a análise desta fonte

    Google Threat Intelligence Group (GTIG) reported in its AI Threat Tracker (published November 16) that it observed a threat actor using an LLM-based framework called exploit_agent. The framework diffs old and patched code, finds the fixed weakness and tries to build and run an exploit with no human in the loop. GTIG calls this its first observation of fully automated vulnerability research and exploitation. The framework was found on an exposed staging server next to stolen victim data. Material there suggested an interest in French media, entertainment and education sites. Russian-language configuration comments were present, but they do not establish the operator's nationality. This record is based on the news article alone. No web research was done, so no primary GTIG source was checked.

    Technical Details

    • exploit_agent performs patch diffing, then exploit generation and execution (N-day exploitation).
    • No success rate is given, and it is not established that every generated exploit worked.
    • GTIG also saw cybercriminals and suspected espionage actors using open-source AI agents. CAIRN was used against targets that included the US aerospace sector. PentAGI was deployed in the wild. CyberStrikeAI and other frameworks were also tested.
    • Other findings: AI-assisted influence operations, including a Gemini API-based platform meant to manage more than 100,000 accounts. China-aligned actors used LLMs to process stolen data. North Korea-aligned activity included AI personas targeting crypto businesses and attempts to manipulate coding assistants' configuration.
    • Google disabled the associated accounts.

    Remediation

    • Patch promptly, especially internet-facing systems.
    • Restrict access to applications that cannot be patched yet.
    • Hunt for prior compromise, because patching does not evict an existing intruder.
  6. thehackernews.com ↗ · pub. 16/11/2026 11:00 · coletado 16/11 15:00

    Google GTIG: Threat Actors Use Agentic AI Frameworks and LLM Patch Diffing to Weaponize N-Day Flaws, Incl. PaperCut NG/MF

    ler a análise desta fonte

    Google Threat Intelligence Group (GTIG) reports that adversarial AI use has moved past manual prompting to autonomous multi-agent offensive frameworks, including ARTEX, CAIRN, CyberStrikeAI and PentAGI, which automate phases of live intrusions. Threat actors are also using LLMs to automate patch reverse engineering through differential analysis, which shortens the time between patch disclosure and exploitation. GTIG cites the PaperCut NG/MF flaws in September as an observed case. Financially motivated actors and suspected China-aligned espionage actors have been seen experimenting with these tools. GTIG also describes LLMjacking, in which stolen LLM credentials and cloud resources are resold through proxy services such as APIMart and JetNight. It also describes the Chinese-language PhaaS platform Jieyou, which has an LLM chatbot that lets low-skilled affiliates customize phishing panels and lures.

    Technical Details

    • Other tools named: ZeroHollow (an inference service for exploit development and LLM routing) and Agent Zero (an open source agent framework with autonomous code execution).
    • LLMs compare source code versions, including pre-release channels, and read release notes and PoC code to weaponize patched n-day flaws.
    • The article gives no CVE IDs, IOCs or named threat actors, and it is based on a GTIG report shared with the publication. No vendor advisory was consulted.
  7. theregister.com ↗ · pub. 16/11/2026 11:00 · coletado 16/11 12:00

    GTIG: Suspected Russian-speaking operator used autonomous AI agent 'exploit_agent' to weaponize patches and steal data from French targets

    ler a análise desta fonte

    The Google Threat Intelligence Group (GTIG) reports in its Q4 2026 AI threat tracker what it calls the first case of a suspected Russian-speaking operator using an AI agent to turn security patches into working exploits with no human in the loop. GTIG found an exposed staging server hosting a multi-agent vulnerability research and weaponization framework, exploit_agent. The framework integrates with several frontier and local models. It continuously diffed popular web application code before and after security updates, identified flaws, developed exploits and executed them automatically. The server also held data stolen from victims. Victim telemetry points to targeting of French media, entertainment and education websites. This summary is based only on the article. No primary Google report or CVE list was consulted.

    Technical Details

    • Post-compromise scripts on the server covered internal reconnaissance, brute-forcing, webshell installation and removal of other attackers' webshells.
    • The tooling used Russian-language content.
    • GTIG also notes other autonomous offensive frameworks in use: ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. PentAGI was seen alongside proxy services against the aerospace sector.
    • Financially motivated actors and suspected PRC-nexus espionage actors are experimenting with these frameworks.

    Remediation

    The article gives no specific remediation. The implication is that the window between patch release and exploit shrinks, so defenders should patch internet-facing web applications quickly.

  8. bleepingcomputer.com ↗ · pub. 16/11/2026 11:00 · coletado 16/11 12:00 primeiro registro

    Google GTIG: threat actors adopt agentic AI frameworks (ARTEX, CAIRN, exploit_agent) to automate exploitation; SOMBERMEME via Claude Code

    ler a análise desta fonte

    Google Threat Intelligence Group (GTIG) reports large-scale adoption of agentic AI workflows by financially motivated and state-sponsored actors to automate vulnerability discovery and exploitation. Open-source pentesting frameworks including ARTEX, CAIRN, CyberStrikeAI and HexStrike AI are being abused. The report cites the first observed case of fully automated discovery and exploitation with no human in the loop, a multi-agent framework called exploit_agent found on an exposed staging server. GTIG projects that unsophisticated, near-autonomous attacks will appear within three months and sophisticated, fully autonomous attacks within six. This summary rests on the article alone, because no web research was run, so the primary GTIG report and the FBI advisory were not checked.

    Technical Details

    • exploit_agent integrates with frontier and local LLMs. It analyzes software updates of popular web apps for signs of flaws and then tries to exploit them. It was used against French media, entertainment and education websites. The staging server also held stolen victim data, reconnaissance data, webshells and brute-forcing scripts.
    • CAIRN was used in an attack on the U.S. aerospace sector.
    • UNC6836, with suspected links to "Midnight Neptune", targeted cryptocurrency firms in North America and Europe. After initial compromise, it changed the settings of a modified Claude Code so that launching it downloaded the SOMBERMEME backdoor.
    • The FBI warned in an advisory that multiple actors have weaponized ARTEX since August against the financial, government, technology, defense and healthcare sectors, among others.
    • Targets are internet-exposed endpoints with low-complexity flaws that LLMs can easily exploit.
    • AI-generated PoCs appeared after the PHP2Shell disclosure and burdened defenders.
    • LLMjacking and AI-driven influence operations are also growing.

    Remediation

    The article gives no specific remediation. It notes that the patch window is shrinking, so defenders should patch internet-exposed systems quickly.

Repetiram sem novidade