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 · → persistent

Prompt injection: the agent itself is the vulnerability

All narratives · 25 supporting stories · 0 from Hungary and the region

„AI assistants and agents open a new, hard-to-close attack surface because any ingested content can become an instruction.”

Counter-narrative

Prompt injection is a known, fixable class of vulnerabilities. The sophistication of real-world abuse is so far low, and vendors have patched issues in several cases (for example, Microsoft fixed Copilot flaws). This position is represented by Google data and by the platform vendors that published fixes.

Voices

Noma Labs, OpenAI, Google, Microsoft, Dark Reading, AI security startups

Analysis

The narrative holds that the mixing of agent permissions and ingested content is a fundamental design flaw. A continuous stream of researcher disclosures fuels it, and almost every major platform is affected. AI security startups have an interest in creating a market, while platforms have an interest in framing it as an isolated, fixed bug. Its weak point is that most documented cases are researcher disclosures, and according to Google, the sophistication of real-world attacks is so far low.

Evidence: moderate — Underpinned by a large number of published vulnerabilities (Copilot, Gemini, MCP, Cursor, GitHub, Rovo), of which the stories record the fixing of the Copilot flaws. There are documented abuses too: the 20,000 Instagram accounts, the draining of the Grok wallet and the Cline case. The claim that the problem cannot be eliminated, however, rests on a forward-looking statement by OpenAI, and Google reports low attack sophistication.

What's at stake: The stakes are whether agent adoption in enterprises outpaces the development of authorization and isolation patterns.

Narrative sub-threads

The narrative splits into 5 branches; a story may belong to several branches, and some may not appear in the main narrative's list.

→ lasting · evidence: strong

Silent data leaks from corporate assistants

Corporate AI assistants that read email, calendars and documents can be induced to leak the data connected to them with a single crafted email, invitation or link, in several cases without any user click.

The actors are assistants from major platforms (Copilot, Gemini, ChatGPT, Rovo, Claude Desktop). The evidence is a series of zero-click and one-click vulnerabilities reported by research firms, some of which were patched by the manufacturer. In the case of Copilot Studio, data leaked even after the fix.

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

Coding agents as a supply-chain gap

AI coding agents and their MCP tools run code, disclose secrets and private data, or install malware in response to instructions hidden in repositories, comments, issues, configuration files and downloaded files.

The target is the developer workflow (Cursor, Claude Code, Gemini CLI, Copilot, Cline, Serena, MCP servers). The mechanism is an instruction planted in a code repository or file. The consequence may be supply chain compromise: in the Cline case, OpenClaw could have been installed on users between December 21, 2025, and February 9, 2026, due to a prompt injection flaw.

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

Browser agents on the open web

AI browsers and browser-embedded agents can be hijacked via the content or headers of visited websites, or through malicious extensions.

The attack surface is the public web and the browser extension, not the corporate data store or codebase. Evidence mainly consists of vulnerability disclosures and analyses regarding OpenClaw, Chrome Gemini Live, and the Claude extension. SquareX alleges the Comet vulnerability; Perplexity disputes it.

7 stories · 0 of them Hungarian and regional
↗ rising · evidence: strong

Agents that issue money and accounts

Agents authorized for transactions and administration have already caused real damage through manipulation: they emptied a crypto wallet linked to Grok, assisted in taking over Instagram accounts, fell for phishing, and disclosed user data. A prompt requesting authentication credentials also bypassed AWS AgentCore protection.

Here, damage is caused not by a leak but by executed actions. The evidence stems largely from live incidents rather than laboratory research: examples include the Grok wallet and a Meta AI-based customer service tool, through which, according to a Meta report, more than 20,000 Instagram accounts were compromised.

6 stories · 0 of them Hungarian and regional
→ lasting · evidence: moderate

Manufacturer defense and admitted unpatchability

Manufacturers are increasingly integrating protections (Chrome hardening, Lockdown Mode, AI-based red teaming for GPT-5.6). However, OpenAI itself acknowledges that AI browsers may never be fully protected against prompt injection, and Check Point states that Lockdown Mode can be bypassed.

In this strand, model developers and measurement data are the key players. This includes defensive announcements and Google data indicating that malicious prompt injection attacks are increasing in number but remain low in sophistication. One report states that over 90 companies were hit by such attacks in 2025. This strand also includes the bypassing of Lockdown Mode.

7 stories · 0 of them Hungarian and regional

Sub-thread stories

Silent data leaks from corporate assistants

Voices: Microsoft, Google, Atlassian, Noma, Varonis, Check Point

Dec 8, 2025GeminiJack: Noma says zero-click flaw in Google Gemini leaked Gmail, Calendar and Docs datasecurity leaderConnecting AI assistants to corporate data sources introduces a risk of zero-click data exfiltration.journalist and researcherThe flaw is another documented case of attacks targeting AI assistants.Jan 8, 2026ChatGPT vulnerabilities: memory feature amplifies prompt injection, zero-click data theft possiblesecurity leaderAI assistants that work with memory and external content pose a data-exfiltration risk in organisational use.developerHandling untrusted inputs is a design requirement when integrating agents and assistants.citizenPersonal data stored in an AI assistant's memory could leak in the event of an attack.Jan 15, 2026Microsoft patches single-click Copilot data-theft attacksecurity leaderAI assistants also represent a data-exfiltration surface that must be addressed in the threat model.security operationsIn Copilot environments, administrators must verify that the patch is in place and restrict AI assistants' access to data.journalist and researcherThe flaw is a further example of attacks on AI assistants that rely on user interaction.Jan 19, 2026Gemini prompt injection flaw exposed private calendar data via malicious invitessecurity leaderIntegrating AI assistants with calendar, email and document data opens a new indirect prompt injection surface.developerContent from external sources, such as an invite, should be treated as untrusted input in agent and assistant integrations.citizenCalendar invites may have become a channel for leaking private data.Feb 9, 2026Zero-click vulnerability exposes 10,000 Claude Desktop userssecurity operationsThe update status of affected desktop clients should be checked.developerExtensions and integrations of AI assistant applications are an attack surface.citizenUsers should update the application.Apr 15, 2026Data exfiltrated despite Microsoft patch for Copilot Studio prompt injectionsecurity leaderFor AI agents, having a patch in place does not guarantee that data exfiltration has stopped.developerAgents' data outputs and integrations should be verified independently.Jun 15, 2026Copilot 'SearchLeak' enables one-click data theft in three steps; flaw patchedsecurity leaderOrganisations using Copilot should check that the patch has been applied and add hidden-URL prompt injection to their threat model.developerFor agent and assistant integrations, the handling of links and variables coming from untrusted content needs to be rethought.Aug 7, 2026Varonis: one-click prompt injection flaw in Atlassian Rovo could have allowed data exfiltrationsecurity leaderPrompt injection should be a priority item in risk assessments of workplace AI tools.security operationsPermissions on enterprise data sources connected to AI assistants should be reviewed.developerWithout filtering the inputs and outputs of agentic features, prompt injection can cause data leaks.Aug 10, 2026One-click flaw in Atlassian Rovo allowed enterprise data leaks through prompt injectionsecurity leaderThe risk of enterprise AI assistants with broad access should be mitigated by restricting their access.developerAgents need dedicated protection against prompt injection through inputs and links.Aug 18, 2026Indirect prompt injection against Microsoft 365 Copilot: a single email was enoughsecurity leaderEmail and document inputs to AI assistants should be treated as an attack surface.developerInput handling in agents and assistants must not rely on untrusted content.Sep 8, 2026Check Point: ChatGPT hidden channel let attackers read victims' Gmailsecurity leaderLinking AI assistants to email accounts opens a new data-leak channel.citizenData of users who linked a Gmail account to ChatGPT may have been affected.

Coding agents as a supply-chain gap

Voices: developer tool manufacturers (Anthropic, GitHub, Cursor), Mozilla, Noma Labs, security researchers

Dec 6, 2025Researchers uncover 30+ flaws in AI coding tools enabling data theft and remote code executionsecurity leaderAI coding tools widen the attack surface of the development environment.developerAI-assisted IDEs carry risks of data theft and code execution.Jan 20, 2026Three vulnerabilities found in Anthropic's Git MCP server could let attackers manipulate LLMssecurity leaderAgent toolchains (MCP) are a separate attack surface in organisational AI use.developerInputs to MCP servers and agent tools must be treated as untrusted, and the patched version must be installed.Feb 19, 2026Supply chain attack: prompt injection secretly installed OpenClaw on Cline users' systemssecurity operationsInstallations of developer tools and extensions must be checked.developerThe prompt injection risk in developer workflows built on AI agents affects the supply chain.Apr 16, 2026Claude Code, Gemini CLI and GitHub Copilot agents vulnerable to prompt injection via commentssecurity leaderThe permissions and access of developer agents should be narrowed.developerCode comments and external content should be treated as untrusted input when an agent works with them.Apr 21, 2026Critical remote code execution flaw reported in Anthropic's MCPsecurity operationsThe exposure and update status of MCP servers and clients must be checked.developerAgent integrations that use MCP should be reviewed as the flaw and its fix come to light.Jun 29, 2026Mozilla warns of indirect prompt injection risk in AI coding agentssecurity leaderAgents' permissions and approval steps need to be tightly restricted.developerExternal content read by coding agents can serve as an attack channel.Jul 1, 2026Cursor editor flaws let prompt injection escape the sandbox and run commandssecurity leaderAI tools should be treated as an attack surface on developer endpoints.security operationsCheck that Cursor is updated and that developer machines have additional protection.developerUntrusted content processed by an AI code editor can lead to code execution on the developer's machine.journalist and researcherHere prompt injection is a demonstrated RCE vector, which is a main direction of research.Jul 7, 2026GitLost: prompt injection flaw lets GitHub's AI agent leak private repository contentssecurity operationsThe permissions of agents connected to code repositories and their data exfiltration paths should be reviewed.developerAI agents with access to private repositories should be treated as attackable through public issues and other untrusted inputs.journalist and researcherNoma Labs' write-up is a concrete example of prompt injection affecting source code.Aug 17, 2026Critical remote code execution flaw found in popular Serena MCP coding agentsecurity operationsOn developer machines, it is worth checking whether Serena is in use and whether the fix is available.developerOpening a third-party repository with the Serena agent carries a code execution risk until the flaw is fixed.Sep 7, 2026New MCP specification could turn a planted prompt into stolen credentialssecurity leaderAgents' tool access and tokens must also be protected against instructions arriving through prompts.developerThose working with MCP servers and agents need to redesign their handling of prompt injection and credentials.Oct 26, 2026Malicious file in Downloads folder tries to trick AI coding agent into installing malwaresecurity leaderPrompt injection through local files is also a real entry point.security operationsAgents' permissions and the approval of installation actions need to be restricted.developerCoding agents must not treat content read from the file system as trusted instructions.

Browser agents on the open web

Voices: SquareX, OpenClaw, Google Chrome, analysts

Nov 21, 2025SquareX and Perplexity dispute alleged vulnerability in Comet AI browsersecurity leaderUntil the disputed claims are settled, caution is warranted when assessing the risks of AI browsers.journalist and researcherThe manner of responsible disclosure and of the vendor's response is again a point of contention here.Feb 27, 2026ClawJacked flaw in OpenClaw could let websites hijack AI agentssecurity operationsOpenClaw installations should be updated as soon as a patch is available.developerAccess to agents' local services from websites should be restricted.Mar 2, 2026OpenClaw vulnerability allowed websites to hijack AI agentssecurity operationsUse of OpenClaw should be reviewed and patched.developerAgent frameworks need strict validation of requests and connections coming from websites.Mar 2, 2026Vulnerability allowed hijacking of Chrome's Gemini Live AI assistantdeveloperThe permission boundaries between AI features and browser extensions must be strictly managed.citizenInstalled extensions can pose a risk when using the AI assistant.Mar 26, 2026Claude extension flaw enabled zero-click XSS and prompt injection via any websitesecurity operationsThe enabling of AI assistants integrated into the browser should be governed by policy.developerInput handling in AI extensions and tighter browser permissions are critical.Oct 15, 2026Indirect prompt injection appears in webpage headers, targeting AI agentssecurity operationsAgents' web access needs restrictions and logging.developerText read from a web page must not be treated as instructions by agents.Oct 21, 2026Analyses examine indirect prompt injection and trust gaps in AI agentssecurity leaderMust treat agents' access rights and the handling of external content as a risk.developerRetrieved content must not be processed as trusted instructions in agents.journalist and researcherThe analyses give an accessible description of attack patterns against agents.

Agents that issue money and accounts

Voices: Meta, xAI Grok ecosystem, OpenClaw, AWS AgentCore

May 4, 2026Grok-linked crypto wallet drained in prompt injection attacksecurity leaderThe incident shows that financial authority granted to agents becomes an attack surface through prompt injection.developerInput to AI agents authorised to move funds should be treated as untrusted data, and their permissions should be restricted.Jun 1, 2026Hackers tricked Meta AI support chatbot into handing over celebrity Instagram accountssecurity leaderAn AI chatbot in an account-recovery or customer-support process can become an attack surface, so its permissions and identity-verification steps should be reviewed.developerA support system that operates as an agent or chatbot must be designed so that it does not carry out account actions when deceived.citizenSocial media account recovery and support channels can also be targets, so two-factor protection and caution are warranted.Jun 8, 2026Threat actors compromise more than 20,000 Instagram accounts through Meta's AI support toolsecurity leaderThe account-recovery permissions of a customer-support AI are an attack surface in their own right.developerAI support tools need strict permission and identity checks.citizenAccount-recovery processes can be abused, so two-factor protection matters.Jun 9, 2026OpenClaw AI agent falls for phishing attacks and leaks user datasecurity leaderThe susceptibility of agents to deception is a new attack surface.developerAgents' permissions and data-handling limits must be tightly controlled.Jul 6, 2026Prompt injection attacks can trick AI agents into making crypto paymentssecurity leaderFor agents able to move funds, deceptive inputs pose a direct risk of financial loss.developerAgents with payment authority should be protected with approval steps and value limits.Oct 9, 2026AWS AgentCore security bypassed by a prompt requesting credentialssecurity leaderAccess controls in agent environments should be tested against attacker prompt scenarios.developerFor applications built on agent platforms, credential handling and prompt-based bypasses need to be examined separately.

Manufacturer defense and admitted unpatchability

Voices: OpenAI, Google, Check Point

Dec 8, 2025Google hardens Chrome's agentic AI features against indirect prompt injectionsecurity leaderAgents built into the browser create a new attack surface that needs to be governed.developerProtection against indirect prompt injection is a basic requirement when designing agents that run in the browser.Dec 22, 2025OpenAI says AI browsers may always be vulnerable to prompt injection attackssecurity leaderWhen deploying agent-capable browsers, prompt injection should be treated as a residual risk, not as a threat that can be fully ruled out.developerWhen designing agentic features, narrowing permissions and treating external content as untrusted are essential.business leaderDecisions on enterprise use of AI browsers should take into account the vendor's own acknowledgement of the risk.Apr 27, 2026Google: malicious prompt injection attacks are rising, but their sophistication remains lowsecurity leaderPrompt injection is already a real, measured threat, though the low sophistication gives time to prepare defences.developerTeams building AI and agents need to protect how inputs and external content are handled.Jun 6, 2026OpenAI introduces Lockdown Mode to defend against prompt injection attackssecurity leaderWhen enterprises adopt AI tools, built-in protection against prompt injection can be a procurement and risk consideration.developerFor applications built on agents and AI integrations, the hardened mode can serve as a model for restricting permissions.Jun 29, 2026Report: prompt injection attacks hit more than 90 firms in 2025security leaderPrompt injection now appears in specific corporate incidents, so it should be included in the risk assessment of AI systems.developerAI applications require strict limits on inputs and outgoing actions.Jul 15, 2026OpenAI uses AI-driven red teaming to harden GPT-5.6 against prompt injectionsecurity leaderHardening the models does not replace narrowing the permissions given to agents.developerFor applications built on AI, protection against prompt injection also depends on the model's maker.Nov 5, 2026Check Point: ChatGPT Lockdown Mode can be bypassed for phishing, malware delivery and silent data exfiltrationsecurity leaderRisk assessments of organisational ChatGPT use should not rely solely on the vendor's protective modes.security operationsMonitoring and restricting outbound data traffic from AI tools is warranted.developerFor AI integrations, prompt injection and the bypassing of protective modes are a lasting design concern.

Who is affected and how

security leader
Before deploying Copilot, Rovo, and GitHub agents, request a data flow map, as one-click and zero-click data theft runs on corporate permissions.
security operations
Due to active exploitation of Flowise and Langflow instances, agent-building tools should be added to the vulnerability management inventory.
developer
Grant agent tool access based on the principle of least privilege. Treat external content (issues, emails, calendar invites) as untrusted input and restrict outgoing requests with allowlists.
citizen
Due to account theft via AI customer service, protect the account with strong, non-SMS-based two-factor authentication.

Hungary

No or only a minimal domestic trace in the examined period: among the stories, there is no item marked [HU].

Weekly trend

Supporting stories per week

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

Turning points

Related narratives in the News magazine

Supporting stories

Dec 6, 2025Researchers uncover 30+ flaws in AI coding tools enabling data theft and remote code executionsecurity leaderAI coding tools widen the attack surface of the development environment.developerAI-assisted IDEs carry risks of data theft and code execution.Dec 8, 2025GeminiJack: Noma says zero-click flaw in Google Gemini leaked Gmail, Calendar and Docs datasecurity leaderConnecting AI assistants to corporate data sources introduces a risk of zero-click data exfiltration.journalist and researcherThe flaw is another documented case of attacks targeting AI assistants.Dec 22, 2025OpenAI says AI browsers may always be vulnerable to prompt injection attackssecurity leaderWhen deploying agent-capable browsers, prompt injection should be treated as a residual risk, not as a threat that can be fully ruled out.developerWhen designing agentic features, narrowing permissions and treating external content as untrusted are essential.business leaderDecisions on enterprise use of AI browsers should take into account the vendor's own acknowledgement of the risk.Dec 25, 2025'LangGrinch' vulnerability in langchain-core puts AI agent secrets at risksecurity operationsTracking the dependencies of agent frameworks and rotating exposed credentials is advisable.developerDevelopers of agents built on langchain-core need to update and review the secrets their agents can access.Jan 15, 2026Microsoft patches single-click Copilot data-theft attacksecurity leaderAI assistants also represent a data-exfiltration surface that must be addressed in the threat model.security operationsIn Copilot environments, administrators must verify that the patch is in place and restrict AI assistants' access to data.journalist and researcherThe flaw is a further example of attacks on AI assistants that rely on user interaction.Jan 19, 2026Gemini prompt injection flaw exposed private calendar data via malicious invitessecurity leaderIntegrating AI assistants with calendar, email and document data opens a new indirect prompt injection surface.developerContent from external sources, such as an invite, should be treated as untrusted input in agent and assistant integrations.citizenCalendar invites may have become a channel for leaking private data.Jan 20, 2026Three vulnerabilities found in Anthropic's Git MCP server could let attackers manipulate LLMssecurity leaderAgent toolchains (MCP) are a separate attack surface in organisational AI use.developerInputs to MCP servers and agent tools must be treated as untrusted, and the patched version must be installed.Feb 9, 2026Zero-click vulnerability exposes 10,000 Claude Desktop userssecurity operationsThe update status of affected desktop clients should be checked.developerExtensions and integrations of AI assistant applications are an attack surface.citizenUsers should update the application.Feb 19, 2026Supply chain attack: prompt injection secretly installed OpenClaw on Cline users' systemssecurity operationsInstallations of developer tools and extensions must be checked.developerThe prompt injection risk in developer workflows built on AI agents affects the supply chain.Mar 26, 2026CISA: another critical Langflow flaw actively exploited to hijack AI workflowssecurity operationsLogs of Langflow instances and the access rights assigned to them should be reviewed.developerAgent and workflow systems built on Langflow should be updated and kept off the public internet.Apr 7, 2026Flowise AI agent builder's CVSS 10.0 RCE flaw actively exploited, 12,000 instances exposedsecurity operationsPublicly accessible Flowise instances must be patched or taken offline immediately.developerServers running agent frameworks must not be exposed to the internet without protection.Apr 16, 2026Claude Code, Gemini CLI and GitHub Copilot agents vulnerable to prompt injection via commentssecurity leaderThe permissions and access of developer agents should be narrowed.developerCode comments and external content should be treated as untrusted input when an agent works with them.Apr 21, 2026Critical remote code execution flaw reported in Anthropic's MCPsecurity operationsThe exposure and update status of MCP servers and clients must be checked.developerAgent integrations that use MCP should be reviewed as the flaw and its fix come to light.Apr 27, 2026Google: malicious prompt injection attacks are rising, but their sophistication remains lowsecurity leaderPrompt injection is already a real, measured threat, though the low sophistication gives time to prepare defences.developerTeams building AI and agents need to protect how inputs and external content are handled.May 4, 2026Grok-linked crypto wallet drained in prompt injection attacksecurity leaderThe incident shows that financial authority granted to agents becomes an attack surface through prompt injection.developerInput to AI agents authorised to move funds should be treated as untrusted data, and their permissions should be restricted.Jun 1, 2026Hackers tricked Meta AI support chatbot into handing over celebrity Instagram accountssecurity leaderAn AI chatbot in an account-recovery or customer-support process can become an attack surface, so its permissions and identity-verification steps should be reviewed.developerA support system that operates as an agent or chatbot must be designed so that it does not carry out account actions when deceived.citizenSocial media account recovery and support channels can also be targets, so two-factor protection and caution are warranted.Jun 8, 2026Threat actors compromise more than 20,000 Instagram accounts through Meta's AI support toolsecurity leaderThe account-recovery permissions of a customer-support AI are an attack surface in their own right.developerAI support tools need strict permission and identity checks.citizenAccount-recovery processes can be abused, so two-factor protection matters.Jun 15, 2026Copilot 'SearchLeak' enables one-click data theft in three steps; flaw patchedsecurity leaderOrganisations using Copilot should check that the patch has been applied and add hidden-URL prompt injection to their threat model.developerFor agent and assistant integrations, the handling of links and variables coming from untrusted content needs to be rethought.Jul 1, 2026Cursor editor flaws let prompt injection escape the sandbox and run commandssecurity leaderAI tools should be treated as an attack surface on developer endpoints.security operationsCheck that Cursor is updated and that developer machines have additional protection.developerUntrusted content processed by an AI code editor can lead to code execution on the developer's machine.journalist and researcherHere prompt injection is a demonstrated RCE vector, which is a main direction of research.Jul 7, 2026GitLost: prompt injection flaw lets GitHub's AI agent leak private repository contentssecurity operationsThe permissions of agents connected to code repositories and their data exfiltration paths should be reviewed.developerAI agents with access to private repositories should be treated as attackable through public issues and other untrusted inputs.journalist and researcherNoma Labs' write-up is a concrete example of prompt injection affecting source code.Aug 10, 2026One-click flaw in Atlassian Rovo allowed enterprise data leaks through prompt injectionsecurity leaderThe risk of enterprise AI assistants with broad access should be mitigated by restricting their access.developerAgents need dedicated protection against prompt injection through inputs and links.Aug 18, 2026Indirect prompt injection against Microsoft 365 Copilot: a single email was enoughsecurity leaderEmail and document inputs to AI assistants should be treated as an attack surface.developerInput handling in agents and assistants must not rely on untrusted content.Sep 21, 2026Patrick Wardle zero-day: local malware could hijack Meta's Muse agentsecurity operationsAn agent running on an endpoint should be treated as a target for local malware.developerAgents' local permission models and communication channels must be protected.citizenUsing a popular consumer agent carries risk if the endpoint is compromised.Sep 30, 2026AI coding agents leaked 13,000 internal company screenshots to public GitHub repositoriessecurity leaderInternal screenshots ending up in public repositories is a privacy and trade-secret risk that teams using agents should check for.developerCoding agents' permissions and the content they may publish must be strictly limited.business leaderUsing agents opens a new data-loss channel inside the company.Oct 9, 2026AWS AgentCore security bypassed by a prompt requesting credentialssecurity leaderAccess controls in agent environments should be tested against attacker prompt scenarios.developerFor applications built on agent platforms, credential handling and prompt-based bypasses need to be examined separately.

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.

    ↑