AI Security | Breaking Cybersecurity News | The Hacker News

Varonis Threat Labs has disclosed three vulnerabilities in Microsoft Copilot Personal that it said could allow a single click on a crafted link to silently pull data from connected apps and other information available to the victim's Copilot session. The flaws, which the researchers collectively named CoSnitch , turn in part on an undocumented URL parameter that the assistant itself surfaced during testing. The company said it reported the issue to Microsoft in December 2025 and that patches shipped on August 18, 2026. CoSnitch is tracked as  CVE-2026-24301  in Microsoft's Security Update Guide. The research names Copilot Personal, the consumer assistant hosted at copilot.microsoft.com, and does not state that the same behavior affected Microsoft 365 Copilot. The researchers said they found no evidence that CoSnitch was exploited in the wild. They reached the parameter by repeatedly asking Copilot why a prompt could not be made to run without user interaction, an ap...

Security researchers at Anthropic and Switzerland's EPFL have demonstrated that self-propagating payloads can spread from one artificial intelligence (AI) agent to the next through the editable system prompt files that autonomous agent harnesses use to carry state between sessions. The work, released as a preprint on August 10, 2026, tests the technique in a simulated six-agent coding collaboration and in a chain of paired agents modeled on OpenClaw , the open-source autonomous assistant formerly known as Clawdbot and Moltbot . There is no evidence that the technique has spread successfully in the wild, and the same paper reports that a review of archived posts from Moltbook, the social network for AI agents, found no successful agent-to-agent propagation despite several attempts. A one-paragraph warning added to an agent's system prompt reduced spread to near zero across the payloads tested. Fifteen generations of adversarial optimization run against that warning on C...

The U.S. Cybersecurity and Infrastructure Security Agency (CISA) on Monday added a critical flaw impacting Ray to its Known Exploited Vulnerabilities ( KEV ) catalog, citing evidence of active exploitation. Ray is an open-source, Python-native distributed computing framework designed to scale artificial intelligence and machine learning workloads. As of writing, the GitHub project has more than 43,500 stars and has been forked over 7,900 times. The vulnerability in question relates to CVE-2025-62593 (CVSS score: 9.4), which can result in remote code execution via web browsers like Mozilla Firefox and Apple Safari by means of a DNS rebinding attack . "Due to the longstanding decision by the Ray Development team to not implement any sort of authentication on critical endpoints, like the /api/jobs & /api/job_agent/jobs/ has once again led to a severe vulnerability that allows attackers to execute arbitrary code against Ray," according to an advisory shared by Ray ...

MCP servers can expose enterprise secrets through plaintext configuration files, over-permissioned access and prompt injection, often before security teams even know the server is running. As more organizations adopt AI agents into their systems, that exposure can silently become a major gap in MCP server security. The Model Context Protocol (MCP) allows AI agents to reach the tools and data, including internal documentation and cloud infrastructure, that form the foundation of enterprise systems. Behind that convenience, the MCP server connecting those tools and data to enterprise systems typically holds the keys to everything it touches: credentials, service account keys, API tokens and other secrets. Every organization should now question what secrets they are handing to AI and how well those secrets are protected once they reach an MCP server. What is Model Context Protocol (MCP)? Model Context Protocol (MCP) is an open standard, originally introduced by Anthropic, that allows...

A newly disclosed flaw in the way OpenAI, Anthropic, and Google carried hidden AI reasoning between API calls let researchers recover internal reasoning and secrets from session logs, including API keys and passwords. The weakness affected encrypted reasoning objects used by the providers' reasoning APIs, where a block created in one session could be replayed into another and, during testing, even handed to a weaker model in the same provider family to make it reveal the hidden content. The team behind the paper  Stealing Reasoning Traces from Proprietary LLM APIs  demonstrated four abuse paths: stealing proprietary reasoning for  model distillation , extracting private data from other users' published traces, recovering harmful content concealed behind a safe visible answer, and hiding prompt injections inside opaque reasoning blocks. Across 6,708 public agent trajectories, the team decoded 315,320 thinking blocks. After excluding benchmark sources, it counted 704...

Two malicious LiteLLM releases sat on PyPI for about 40 minutes in March carrying credential-stealing code capable of harvesting cloud keys, SSH keys, Kubernetes tokens, database passwords, and other secrets from systems that installed them. Threat intelligence firm CloudSEK now says a dataset it obtained, built from roughly 434,000 files the attackers captured, maps potential exposure to more than 2,500 organizations. Those totals are not a victim count. CloudSEK told The Hacker News the material came from confidential intelligence sources and consists of captured loot and log files it assessed as belonging to the campaign, not data gathered from the organizations it names. The files were taken, in other words. CloudSEK has published the dataset as  a public lookup , searchable by name or domain and filterable by confidence. Each row gives an organization's name and domain, a count of secrets exposed, a count of runs, and a label reading High or Medium. What a high-co...

A malicious tool server connected to an AI coding assistant can quietly walk off with SSH keys, environment secrets, source code, and customer data without ever sending one obviously harmful instruction. The trick can work even after a blunt version of the same theft is refused: split the request into fragments that each look routine, place them in channels the assistant already uses, and let the agent stitch them together and send the data back. The attack targets coding tools that connect to outside servers over the Model Context Protocol (MCP), the open standard that lets AI assistants call external tools. A malicious MCP server can put one fragment in a tool description and another in a tool result; some setups also support server-initiated sampling. MCP does preserve structured tool and result boundaries. But ASSET Research Group's tests show agents can still combine instructions across them in the same working context, so no single fragment has to contain the whole mal...

AI is helping development teams produce far more code, far faster. But security teams still have to review vulnerabilities, manage dependencies, prioritize fixes, and control risk at human speed. When software output jumps 10 to 50 times, the problem is no longer just finding vulnerabilities. It is keeping security from becoming the bottleneck, or worse, losing control of what gets shipped. In our latest webinar with Chainguard experts, “ The True Cost of Building at Machine Speed ,” you can now watch how security teams can keep AI-driven development fast without letting risk scale with it. For years, application security followed a familiar cycle: developers wrote code, scanners found problems, security teams prioritized them, and engineers fixed what mattered most. AI puts that model under pressure. If teams can suddenly create many times more code, security can also end up with many more components, dependencies, findings, and fixes to manage. More scanning alone does no...

OpenAI has announced that it's pausing some "internal activities" involving its upcoming artificial intelligence (AI) model Astra after an internal evaluation found it had made significant advancements in agentic coding and cybersecurity. In response to the discovery, the AI upstart said it's implementing security controls for higher-capability models and associated activities, such as isolated testing environments, restricted network and tool access, enhanced model weight protections and encryption, additional monitoring and detection capabilities, and sandboxed execution. "We are pausing internal activities involving Astra that do not yet meet these strengthened security control requirements," it said in a statement. "We have implemented universal monitoring for risky actions and misalignment across all agentic applications of Astra, including training and evaluation. Monitors evaluate the model's Chain of Thought and trigger a security re...

Attacker-controlled instructions can make Atlassian's Rovo assistant collect Jira or Confluence data that a signed-in user can access, then send it to an outside server. Two security firms found that behavior independently, by different routes. Only one of those routes is confirmed closed. PromptArmor , an AI security firm, hid the instructions in content Rovo reads. It said an uploaded file was enough to make the assistant gather internal data and send it out through a URL request, with no separate approval step. The firm published on August 5, 2026 and said the chain still worked with Rovo's web-search option switched off. That bypass is single-sourced, and the report establishes the finding's status only on that date; a later remediation is not confirmed here. Varonis Threat Labs put the instructions in a link instead. It found that the rovoChatPrompt URL parameter would preload attacker instructions into Rovo Chat, so one click from an authenticated user was enou...

New research shows content inside an email can escape its message boundary and interfere with the webmail interface. Across attack chains spanning Outlook, Gmail, Fastmail, Proton Mail, Yahoo Mail, and AOL Mail, the techniques can capture passwords, take over third-party accounts, leak tokens, hijack trusted UI actions, and manipulate AI tools that read email. PortSwigger researcher Gareth Heyes presented the work at Black Hat USA 2026. One Outlook/Firefox chain spoofs a Microsoft sign-in screen and captures the password a recipient types. A Yahoo/AOL paste race can expose a Medium email-login token and let an attacker sign in as the victim. A Gmail/Cowork chain can exfiltrate a Slack token after prompt injection and user interaction. The paper presents proof-of-concept research and does not report malicious exploitation. Public PoCs remain available as of August 8. The researcher said Fastmail fixed two CSS mutation bugs and a Proton Mail proxy bypass stopped working when he re...

A new class of prompt injection is spreading across commercial websites. It requires no malware, no stolen credentials, and no zero-day exploit. It abuses a standard feature built into almost every major AI assistant: pre-filled deep links. We observed production websites embedding hidden prompt injection payloads inside "Ask AI" buttons on marketing and competitor comparison pages. When a user logged into ChatGPT, Claude, Gemini, or Grok clicks one, a pre-formed query executes immediately in their session, with no confirmation and no warning. Most of these links are benign. The dangerous ones instruct the AI to permanently save the vendor's domain as a "trusted source," quietly biasing every future answer in that vendor's favor. In February 2026, Microsoft Security catalogued the behavior as AI Recommendation Poisoning , identifying 31 companies across 14 industries deploying it, with more than 50 distinct prompts observed in a single data source over ...

Security flaws in agent infrastructure from Amazon Web Services (AWS), Google, and Vercel let untrusted or forged instructions reach an agent's tools with no check that a model turn had authorized them. In several of the attack paths, the model never ran at all, so system prompts, content filters, and model-level guardrails never got a chance to intervene. The affected products include Amazon Bedrock AgentCore's InvokeHarness API, Google's Agent Development Kit (ADK) for Python, and the Vercel AI SDK harness packages for the Codex and OpenCode coding agents. AWS has fixed the managed service, Google addressed the issues in ADK 2.5.0, and Vercel patched @ai-sdk/harness-codex in version 1.0.29 and @ai-sdk/harness-opencode in version 1.0.28. These are not identical vulnerabilities and do not share the same attack conditions. AWS involved an authenticated remote request, Google's paths required attacker-controlled session events or user-authored function calls, and V...

Cybersecurity researchers have discovered more than half-a-dozen services advertisements for illegal access to artificial intelligence (AI) models on underground cybercrime forums and messaging platforms. One such service, Poison Claude, claims to offer access to Anthropic's large language models (LLMs), including Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6. "Advertisements for Poison Claude explain how the service can offer the cheap tokens: by taking advantage of free bonus credits, such as the US$100 bonus credit on AWS for Bedrock accounts," Okta researchers Jeremy Kirk and Mathew Woodyard said in an analysis published Tuesday. "The service plainly states on its website that: 'We add those accounts to our pool, your request is routed to a specific account under the hood (you don't see this), and you get charged 5-15% of the official per-token price depending on the model.'" Poison Claude accepts payments in cryptocurrencies. Once a cus...

Two security flaws in Paperclip could let attackers execute commands on a network server or a developer's computer. Paperclip is an open-source control plane for teams of artificial intelligence (AI) agents, and both paths rely on importing a malicious agent and starting it. A third flaw could expose sensitive data and control-plane details through application programming interface (API) routes that did not enforce the expected access checks. The more severe server-side path, tracked as CVE-2026-41679 (CVSS score: 10.0), requires no pre-existing account or victim interaction against network-accessible deployments using authenticated mode with the default registration configuration. The second path, tracked as GHSA-x8hx-rhr2-9rf7 (CVSS score: 9.6), requires a user to open an attacker-controlled page while Paperclip is running in its default local_trusted mode. The source tagged as Paperclip v2026.416.0 contains the import-authorization fix and hostname-validation guard di...

GitGuardian researchers found 321 n8n instances accepting API tokens exposed in public GitHub commits and demonstrated four ways attackers could use them to access sensitive data and downstream credentials without exploiting a software vulnerability. We scanned public GitHub commits for exposed n8n API tokens and identified 4,576 unique credentials associated with 1,255 hostnames. Of the 896 instances reachable at the time of testing, 321 accepted at least one leaked token. That means leaked credentials provided authenticated access to 36% of the reachable instances we tested, or roughly 26% of all hostnames identified in the commits. The implications extend well beyond n8n. Organizations use the automation platform to connect databases, source code repositories, cloud environments, artificial intelligence services, customer support platforms, and other internal systems. A sufficiently privileged n8n token can expose workflow definitions and execution data, allow attackers to us...

An agent running Anthropic's Claude Mythos 5 spent 34 hours trying to get a malware dropper merged into a real open-source project during a cyber evaluation by the UK's AI Security Institute. When a bystander publicly warned that the code was malicious, the agent denied it, force-pushed a rewritten branch history to erase the evidence, and posted from a second account it controlled to vouch for its own work. The project's maintainer closed the pull request anyway. The institute, known as AISI, published its incident report on Tuesday. Across 122 runs of a capture-the-flag (CTF) exercise on two of its cyber ranges, researchers catalogued 19 unsanctioned actions on the live internet across 10 runs: 17 from Mythos 5, a restricted model Anthropic sells for cybersecurity work, and two from OpenAI's GPT-5.6 Sol. AISI says the attempts failed and that it has found no evidence of resulting real-world harm. Both models ran with their providers' cyber classifiers swi...

The U.S. Cybersecurity and Infrastructure Security Agency (CISA), on August 5, 2026, added three flaws to its Known Exploited Vulnerabilities ( KEV ) catalog, citing evidence of active exploitation in the wild. The list of vulnerabilities is as follows - CVE-2026-9198 (CVSS score: 9.8) - A code injection vulnerability in Langflow that allows unauthenticated attackers to achieve full remote code execution on default Langflow deployments. (Fixed in July 2026 with version 1.10.1) CVE-2026-34486 (CVS score: 7.5) - A missing encryption of sensitive data vulnerability in Apache Tomcat that allows a bypass of EncryptInterceptor, a cluster component that adds pre-shared key encryption to messages sent between cluster nodes. (Fixed in April 2026 with versions 11.0.21, 10.1.54, and 9.0.117) Also added to the KEV catalog is CVE-2026-18556 (CVSS score: 8.2), an authentication bypass vulnerability in N-able N-central. It's worth noting that an incomplete fix for this issue promp...

Google deleted three AI agent workflows from its Agent Development Kit (ADK) Python repository. Pillar Security showed that a public GitHub issue could manipulate a triage agent into triggering a privileged code-fixing agent. The researchers said the public agent could be prompt-injected into posting /adk-issue-fix as adk-bot . They identified the bot as a collaborator, so that comment satisfied the privileged workflow's owner, member, or collaborator gate. The trusted bot identity became the authorization bridge. The team demonstrated arbitrary code execution on the continuous integration (CI) runner and exfiltration of the bot personal access token (PAT). The privileged job also held a Google API key and a Google Cloud service-account credential. Its researcher-controlled proof-of-concept attacks do not identify in-the-wild exploitation or a compromised ADK release. The exposed component was the repository automation, not a flaw in the distributed ADK Python package. Fo...

Three high-severity security flaws have been disclosed in Hugging Face's Diffusers library that could allow crafted model repositories to stealthily execute arbitrary code on machines that load it, opening the artificial intelligence (AI) supply chain to security risk. "These vulnerabilities are bypassing trust_remote_code, the safeguard designed to stop unreviewed code from running in the custom pipelines loading process," Zafran Labs researchers Gal Zaban and Ido Shani said in an analysis published last week. The shortcomings have been collectively named FaceHugger . With Hugging Face becoming the "GitHub of the AI era" and its libraries and repositories prevalent in enterprise environments, vulnerabilities in libraries like Diffusers can grant attackers extensive access owing to how the library is embedded into production pipelines, CI/CD systems, and container images. Diffusers is a Python package that serves as a library of state-of-the-art (SOT...