Issue 53 · Nov 16–22, 2026Agent Age News magazine · Other editions · Search · Method · HU / ENG

Incidents and influence attempts in the age of agents

Narrative · ↗ rising

AI is the defenders' bug hunter

All narratives · 10 supporting stories · 1 from Hungary and the region

„Frontier models reportedly find thousands of zero-day vulnerabilities; if defenders receive them first, AI favors defense.”

Counter-narrative

The same capability enables attackers to produce exploits at an industrial scale, while mass flaw discovery causes outages. This is illustrated by governments and companies patching all night due to the PHP flaw.

Voices

Anthropic (Project Glasswing), Mozilla, OpenAI, Google Project Zero, US official, AP

Analysis

According to the narrative, the models' bug-finding capability tips the balance toward defenders if access is rationed in a controlled way. It is fed by lab announcements, and the narrative also justifies keeping the models closed. AI labs have an interest in presenting their capabilities as defensive tools. Its weak point is patching capacity. According to Anthropic, discovery has accelerated so much that disclosure deadlines need to be shortened, and the PHP case showed the burden of patching. Google already writes that attackers are using AI agents to turn patches into exploits.

Evidence: moderate — The 22 security bugs fixed by Mozilla and the PHP outages are measured facts. The “thousands of zero-days”, however, is the developer's own statement, the claim about classified systems comes from an unnamed official, and the 948 exploits are a figure from a single statement.

What's at stake: The stake is whether AI bug hunting increases defenders' advantage or, by overloading repair capacity, attackers'.

Narrative sub-threads

The narrative splits into 4 branches; a story may belong to several branches, and some may not appear in the main narrative's list.

→ lasting · evidence: moderate

bug-hunting results from frontier models

Developers, their partners and researchers say AI models and agents are finding large numbers of severe bugs in real-world software. Examples include OpenSSL, Firefox, classified US government systems, Android and Java. An OpenAI AI pipeline generated 948 Linux privilege-escalation exploits in 30 hours.

This branch rests on the fact of discovery: it reports manufacturer and partner figures (500+, 22, 271, 948) and does not address the distribution of access or the disclosure system.

11 stories · 0 of them Hungarian and regional
→ lasting · evidence: strong

Early access for vetted defenders

AI labs and states release cyber-capable models through controlled programs. Government bodies, EU institutions, open-source maintainers and a narrow circle of companies receive them first. The aim is to secure an advantage for defenders.

Here the mechanism is distribution and gatekeeping: Trusted Access, Daybreak, Gold Eagle, ENISA–Mythos. The actors are governments and agencies, not the vulnerabilities themselves.

11 stories · 1 of them Hungarian and regional
↗ rising · evidence: strong

Overload of the disclosure and patching process

The vulnerability management system is under strain. CVE submissions have increased by 263 percent since 2020, prompting NIST to switch to risk-based NVD processing. AI-slop reports are contaminating the CVE process with false vulnerabilities. Citing AI-accelerated discovery, Anthropic and Google Project Zero are shortening the disclosure deadline to 30 days by 2028. A PHP vulnerability discovered by AI caused government websites in several countries to go offline, including magyarorszag.hu.

This branch shows the institutional consequence: NIST, CVE Root, Project Zero, and Anthropic deadlines, plus an actual shutdown. The actors are owners of vulnerability management systems and operators.

8 stories · 1 of them Hungarian and regional
↗ rising · evidence: moderate

Who will get it first: an attacker race

That defenders gain the upper hand is not guaranteed. Cybercriminal groups and Chinese hackers also use AI to find zero-day bugs, attackers use AI agents to turn patches into exploits, and the cost of zero-day hunting could drop to a few dollars. However, according to a VulnCheck researcher, only 1% of AI-discovered vulnerabilities have so far been exploited in live attacks.

This branch examines the narrative's premise from the attacker's side: it relies on Google, Amazon, and Anthropic observations, as well as VulnCheck data, not on the labs' own defensive results.

7 stories · 0 of them Hungarian and regional

Sub-thread stories

bug-hunting results from frontier models

Voices: Anthropic, Mozilla, OpenAI, Mandiant

Feb 5, 2026Anthropic says Claude Opus 4.6 validated 500+ high-severity vulnerabilitiessecurity leaderAI-based vulnerability discovery can be built into defensive workflows, but the vendor's claim needs independent verification.developerAI support can be effective in code review, but validating the findings still requires human work.journalist and researcherThe 500+ figure is a vendor claim; the methodology awaits independent confirmation.Feb 5, 2026AI discovers 12 vulnerabilities in OpenSSLsecurity operationsFor systems that use OpenSSL, track the patches that are released.developerAI-based code analysis can produce results even in critical open-source libraries.journalist and researcherThe results of AI-driven vulnerability discovery can be checked against the bug reports.Mar 6, 2026Anthropic's Claude finds 22 security vulnerabilities in Firefoxsecurity leaderOn the defensive side, using AI to uncover flaws can be built into vulnerability management.developerAI-based bug hunting offers a new way to audit large codebases.Apr 7, 2026Anthropic says Claude Mythos Preview found thousands of zero-days; Project Glasswing launchedsecurity leaderPrepare for a wave of patches and faster discovery of vulnerabilities; the capabilities have not yet been independently verified.security operationsExpect more patches for major systems, so patching processes should be sped up.developerBecause AI can quickly find decades-old hidden bugs, code auditing is becoming more important.public-sector decision-makerThe limited-access release of these capabilities raises questions about regulating AI capabilities and about defensive access to them.journalist and researcherClaims about the capabilities rest on Anthropic's own statements and need independent verification.Apr 21, 2026Mozilla says Anthropic's Mythos found 271 security vulnerabilities in Firefox 150security leaderUsing AI defensively in vulnerability discovery can speed up the patching cycle.developerAI-based vulnerability discovery can produce a large volume of findings, and handling them burdens the development process.Jun 24, 2026US official says Anthropic's Mythos model found vulnerabilities in classified government systemssecurity leaderAI-based vulnerability discovery can be a real tool on the defensive side too, including in classified environments.public-sector decision-makerFlaws found in government systems raise the question of how AI tools should be introduced under regulation.Aug 17, 2026Wiz AI agent finds Snowflake flaw that GitHub Copilot missedsecurity leaderAgent-based offensive testing can complement traditional assessment.developerCode-assistant review is no substitute for targeted security testing.Aug 18, 2026Mandiant's agentic vulnerability discovery system finds over 100 critical flaws in two dayssecurity leaderAI-based vulnerability discovery can speed up detection, but remediation capacity must grow too.developerAutomated discovery may bring a larger volume of bug reports.Sep 28, 2026Open-source AI security agent finds 24 Android vulnerabilitiesdeveloperAI-based vulnerability scanning can provide a practical tool for code review.journalist and researcherThe method and results are measurable examples of defensive use of agents.Nov 5, 2026OpenAI AI pipeline generates 948 Linux root privilege-escalation exploits in 30 hourssecurity leaderVulnerability management must account for the time available before an exploit appears, which AI may shorten further.security operationsPatching Linux systems quickly and detecting privilege-escalation attempts matter even more.developerAutomated exploit discovery raises the importance of security review in kernel and system-level code.journalist and researcherWithout independent verification of the figure and the methodology, the real impact is hard to judge.Nov 18, 2026AI used to uncover critical Java vulnerability, Euronews reportsdeveloperFor Java-based products, watch for the fix to this flaw.journalist and researcherAnother example of AI-based vulnerability research.

Early access for vetted defenders

Voices: OpenAI, Anthropic, White House, ENISA

Apr 14, 2026OpenAI expands Trusted Access for Cyber and offers GPT-5.4-Cyber to vetted defenderssecurity leaderAs AI capabilities grow, defenders may also need access programs.security operationsThe model, built for defense, is available through vetted access.Apr 16, 2026White House to give US agencies access to Anthropic's Mythossecurity leaderGovernment access shows that state defence, too, is counting on the model's capabilities.public-sector decision-makerUS federal bodies are getting direct access to the model.Jun 1, 2026Anthropic offers EU access to its Mythos hacking technologysecurity leaderWider availability of AI-driven vulnerability discovery may change both defensive and threat models.public-sector decision-makerEU institutions must decide how to obtain access to an AI-based offensive capability and on what terms.Jun 17, 2026EU cybersecurity agency to meet Anthropic on Thursday, Commission sayssecurity leaderThe meeting could set the direction for risk-management expectations towards AI providers.public-sector decision-makerTalks between EU regulators and agencies bear on how the security of AI models is handled.Jun 22, 2026OpenAI's Patch the Planet initiative supports open source maintainerssecurity leaderAI-based vulnerability research could speed up the patching of open source dependencies.developerOpen source maintainers may receive AI support in finding and fixing vulnerabilities.Jul 14, 2026White House announces 'Gold Eagle' AI clearinghouse for cyber vulnerabilitiessecurity leaderThe discovery and handling of vulnerabilities may speed up.developerDevelopers of advanced AI may face expectations of cooperation.public-sector decision-makerThe US model shows how early government access to AI capabilities can be organised.Aug 10, 2026OpenAI expands Daybreak defence programme and launches new cybersecurity modelsecurity leaderAI-based defensive tools could add a new layer to SOC work.security operationsThe new model's access terms and effectiveness are worth evaluating.Sep 10, 2026EU cybersecurity agency ENISA granted access to Anthropic's Mythos 5 model, Commission sayspublic-sector decision-makerDirect model access for the EU agency strengthens oversight of AI risks.journalist and researcherThe terms and scope of the access are not clear from the announcement.Sep 22, 2026OpenAI's Astra rollout limits cyber tools to 16 firmssecurity leaderNot everyone has access to the most capable AI-based defensive tools.journalist and researcherThe access restriction raises questions about transparency.Sep 23, 2026OpenAI extends Daybreak cyber defence tools to Ukraine for civilian protectionsecurity leaderDefensive AI tools are becoming available in the affected region.public-sector decision-makerAI capabilities are also entering government cyber defence cooperation.Oct 6, 2026Anthropic launches Cyber Verification Program and Cyber Mission initiativesecurity operationsDefence teams may gain verified access to the cybersecurity capabilities of AI models.journalist and researcherFor researchers, the conditions of the verification program affect how far the models are available for research purposes.

Overload of the disclosure and patching process

Voices: NIST, ENISA, Google Project Zero, Anthropic

Apr 16, 2026NIST moves National Vulnerability Database to risk-based CVE triage amid record submissionssecurity leaderTools and processes should be reviewed for their dependence on the NVD.security operationsVulnerability management that relies on the NVD needs another source for CVEs that have not been enriched.journalist and researcher"Not Scheduled" CVEs do not disappear, but NIST will not enrich them automatically.Apr 20, 2026NIST moves to risk-based NVD model as CVE submissions rise 263% since 2020security leaderOrganisations will increasingly have to prioritise vulnerabilities and score them themselves.security operationsVulnerability management tools can no longer count on NVD enrichment and scores for every CVE.Aug 5, 2026AI slop pollutes the CVE process with fake vulnerabilitiessecurity operationsFake CVE entries add to the burden of prioritisation, so sources need to be verified.developerMaintainers may be swamped by a flood of fake AI-generated reports.journalist and researcherScreening and authenticating vulnerability reports is becoming more important.Aug 10, 2026ENISA expands CVE Root to 20 CNAssecurity operationsCoordinated vulnerability disclosure can rely on a wider European network.developerThe expansion of the CNA network affects how vulnerabilities are reported.Oct 21, 2026magyarorszag.hu and government sites in Slovakia go offline over PHP flawsecurity operationsDeliberately shutting down the government portal and restarting it safely after patching is their task.public-sector decision-makerThe outage affects electronic administration, so decisions are also needed on communication and interim arrangements.citizenWhile the portal is down, online administration is not available there.Oct 21, 2026AI-discovered critical PHP flaw takes government websites offline worldwidesecurity leaderMust immediately assess where PHP runs across the estate and speed up patching.security operationsResponsible for urgent patching and monitoring logs for attack attempts.developerApplications built on PHP must be updated and tested against the patched version.business leaderWebsites and services running PHP carry a risk of outages and attacks.public-sector decision-makerSome government sites were taken offline because of the flaw, so decisions are also needed on the outage and on recovery.Nov 15, 2026Google Project Zero to shorten vulnerability disclosure deadline to 30 days by 2028security operationsPatches and public details will follow each other more quickly.developerVendors have less time to produce a fix before public disclosure.journalist and researcherThe change in public disclosure practice also affects researchers' disclosure timelines.Nov 16, 2026Anthropic and Google Project Zero both shorten vulnerability disclosure deadlinessecurity operationsDetails become public sooner, so a faster patch cycle is needed.developerVendors must release fixes in less time, or the flaw may be disclosed publicly.journalist and researcherChanges in disclosure practice and in the frameworks of coordinated responsible disclosure need to be tracked.

Who will get it first: an attacker race

Voices: Google, Amazon, Anthropic, VulnCheck

Jan 23, 2026Heise reports that AI-generated zero-day exploits are already a realitysecurity leaderAI can speed up the exploitation of vulnerabilities, so patching windows should be shortened.journalist and researcherThe technical basis for the claim should be checked in the source article.May 11, 2026Google detects AI-developed zero-day exploit tied to cybercriminal groupsecurity leaderAI-accelerated vulnerability research and exploit development mean shorter defence windows, which should be factored into threat models and patch priorities.security operationsBecause of the reported bypass of two-factor protection, it is worth reviewing how authentication processes are monitored.developerBecause AI can uncover flaws, code auditing and rapid patching matter even more.journalist and researcherThe case is a defender-documented example of AI-assisted exploit development that researchers should follow.Jul 29, 2026VulnCheck: only 1% of AI-discovered vulnerabilities exploited in the wildsecurity leaderAI-driven vulnerability research may increase the patching workload, while the exploitation rate remains low for now.journalist and researcherThe measured rate adds data to the debate on AI's offensive capabilities.Sep 10, 2026Anthropic: AI lets smaller actors run state-level hacking campaignssecurity leaderThe report says AI appears in attacks at varying levels of autonomy, which affects threat modelling.security operationsAnthropic says attacker sophistication no longer depends on the size of the actor, so detection assumptions need rethinking.journalist and researcherThe report gives details of cases ranging from cyberattacks to influence operations, surveillance, biology and weapons building.Nov 16, 2026AI can now find zero-day flaws for as little as $4, report sayssecurity leaderAs the cost of finding vulnerabilities falls, shortening the time to patch becomes more important.developerDefenders can also use AI tools to scan code for security flaws.Nov 16, 2026Google: attackers use AI agents to turn software patches into working exploitssecurity leaderPatch-management SLAs should be re-evaluated in light of automated exploit development.security operationsThe window between a patch release and an exploit appearing is narrowing, so patching speed matters even more.developerPublished security fixes could be quickly reverse-engineered and exploited by AI agents.Nov 20, 2026Amazon says Chinese-nexus hackers used Cairn AI tool to find apparent zero-day in attack on universitysecurity leaderAI-assisted zero-day discovery narrows defenders' advantage over unknown flaws.journalist and researcherUniversity and research environments can be prime targets for AI-driven espionage.

Who is affected and how

security operations
Following the PHP case pattern, an AI-discovered flaw can cause mass outages overnight. Maintain an up-to-date inventory of runtime environments to identify affected systems within hours.
developer
Prepare for a shorter disclosure window: the 30-day deadline for severe flaws also pressures the patching cycle of your own product. It is advisable to run AI-based code checks before release.
public-sector decision-maker
Government service outages are a realistic scenario: in the magyarorszag.hu case, up to a 48-hour processing suspension had to be anticipated. A pre-agreed communication and alternative processing plan is required.
journalist and researcher
Results from closed programs cannot be independently verified, so seek Mozilla-type figures validated by third parties.

Hungary

The only domestic trace is the shutdown of magyarorszag.hu due to a PHP bug affecting multiple countries, potentially causing a 48-hour service interruption. This indicates that Hungarian state systems are directly impacted by the international bug discovery wave.

Weekly trend

Supporting stories per week

2Nov 17–23: 0Nov 24–30: 0dec. 1–7.: 0dec. 8–14.: 0dec. 15–21.: 0dec. 22–28.: 0Dec 29 – Jan 4: 0jan. 5–11.: 0jan. 12–18.: 0jan. 19–25.: 0Jan 26 – Feb 1: 0febr. 2–8.: 0febr. 9–15.: 0febr. 16–22.: 0Feb 23 – Mar 1: 0Mar 2–8: 1Mar 9–15: 0Mar 16–22: 0Mar 23–29: 0Mar 30 – Apr 5: 0Apr 6–12: 1Apr 13–19: 0Apr 20–26: 0Apr 27 – May 3: 0May 4–10: 0May 11–17: 0May 18–24: 0May 25–31: 0Jun 1–7: 0Jun 8–14: 0Jun 15–21: 0Jun 22–28: 1Jun 29 – Jul 5: 0Jul 6–12: 0Jul 13–19: 0Jul 20–26: 0Jul 27 – Aug 2: 0Aug 3–9: 0Aug 10–16: 0Aug 17–23: 0Aug 24–30: 0Aug 31 – Sep 6: 0Sep 7–13: 1Sep 14–20: 0Sep 21–27: 1Sep 28 – Oct 4: 0Oct 5–11: 0Oct 12–18: 0Oct 19–25: 2Oct 26 – Nov 1: 0Nov 2–8: 1Nov 9–15: 0Nov 16–22: 2Nov 17–23Nov 16–22

Turning points

Related narratives in the News magazine

Supporting stories

Mar 6, 2026Anthropic's Claude finds 22 security vulnerabilities in Firefoxsecurity leaderOn the defensive side, using AI to uncover flaws can be built into vulnerability management.developerAI-based bug hunting offers a new way to audit large codebases.Apr 7, 2026Anthropic says Claude Mythos Preview found thousands of zero-days; Project Glasswing launchedsecurity leaderPrepare for a wave of patches and faster discovery of vulnerabilities; the capabilities have not yet been independently verified.security operationsExpect more patches for major systems, so patching processes should be sped up.developerBecause AI can quickly find decades-old hidden bugs, code auditing is becoming more important.public-sector decision-makerThe limited-access release of these capabilities raises questions about regulating AI capabilities and about defensive access to them.journalist and researcherClaims about the capabilities rest on Anthropic's own statements and need independent verification.Jun 24, 2026US official says Anthropic's Mythos model found vulnerabilities in classified government systemssecurity leaderAI-based vulnerability discovery can be a real tool on the defensive side too, including in classified environments.public-sector decision-makerFlaws found in government systems raise the question of how AI tools should be introduced under regulation.Sep 10, 2026Anthropic: AI lets smaller actors run state-level hacking campaignssecurity leaderThe report says AI appears in attacks at varying levels of autonomy, which affects threat modelling.security operationsAnthropic says attacker sophistication no longer depends on the size of the actor, so detection assumptions need rethinking.journalist and researcherThe report gives details of cases ranging from cyberattacks to influence operations, surveillance, biology and weapons building.Sep 22, 2026AI-guided malware: self-rewriting code, CLOSEDQUORUM and Cisco Talos's CAIRN toolsecurity operationsNew detection approaches are needed for shifting malware that uses AI.journalist and researcherCAIRN is open source, so researchers can try it and extend it.Oct 21, 2026AI-discovered critical PHP flaw takes government websites offline worldwidesecurity leaderMust immediately assess where PHP runs across the estate and speed up patching.security operationsResponsible for urgent patching and monitoring logs for attack attempts.developerApplications built on PHP must be updated and tested against the patched version.business leaderWebsites and services running PHP carry a risk of outages and attacks.public-sector decision-makerSome government sites were taken offline because of the flaw, so decisions are also needed on the outage and on recovery.Oct 21, 2026magyarorszag.hu and government sites in Slovakia go offline over PHP flawsecurity operationsDeliberately shutting down the government portal and restarting it safely after patching is their task.public-sector decision-makerThe outage affects electronic administration, so decisions are also needed on communication and interim arrangements.citizenWhile the portal is down, online administration is not available there.Nov 5, 2026OpenAI AI pipeline generates 948 Linux root privilege-escalation exploits in 30 hourssecurity leaderVulnerability management must account for the time available before an exploit appears, which AI may shorten further.security operationsPatching Linux systems quickly and detecting privilege-escalation attempts matter even more.developerAutomated exploit discovery raises the importance of security review in kernel and system-level code.journalist and researcherWithout independent verification of the figure and the methodology, the real impact is hard to judge.Nov 16, 2026Anthropic and Google Project Zero both shorten vulnerability disclosure deadlinessecurity operationsDetails become public sooner, so a faster patch cycle is needed.developerVendors must release fixes in less time, or the flaw may be disclosed publicly.journalist and researcherChanges in disclosure practice and in the frameworks of coordinated responsible disclosure need to be tracked.Nov 16, 2026Google: attackers use AI agents to turn software patches into working exploitssecurity leaderPatch-management SLAs should be re-evaluated in light of automated exploit development.security operationsThe window between a patch release and an exploit appearing is narrowing, so patching speed matters even more.developerPublished security fixes could be quickly reverse-engineered and exploited by AI agents.

Machine analysis based on story titles and summaries, with a verification pass. The source of the facts is the article linked at each story.

What do you make of this piece?

The split shows after 20 votes.

Comment on LinkedIn
Saved pieces

    Weekly email

    One email a week with the main stories from the Agent Age editions. You confirm by email and can unsubscribe in one click.

    Almost done: we sent you an email. Open it and press the confirm link. If it does not arrive within a few minutes, check your spam folder.

    ↑