AI News Today

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

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

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

A cluster of nearly 800 malicious packages has been published to the npm registry as part of a new campaign designed to deliver cross-platform malware targeting Windows, Mac, and Linux systems. "These packages appear to use AI slop squatted, or randomly generated typo-squatting package names, but all of them deliver a powerful RAT and infostealer payload," OpenSourceMalware researcher Paul

WordPress has fixed a pre-authentication reflected cross-site scripting (XSS) flaw in its login screen that affects every version of the content management system. pwn.ai demonstrated how the flaw can be chained into PHP code execution on the server when a logged-in administrator interacts with an attacker-controlled page. Tracked as CVE-2026-64638 (CVSS score: 8.9), the high-severity

PortSwigger says HTTP Terminator, an artificial intelligence (AI)-assisted research system built by James Kettle, generated and proved new HTTP desynchronization techniques after exploring 30,000 candidate desync vectors. PortSwigger said a separate human-guided discovery cascade also exposed a zero-day in Apache Traffic Server. Kettle said HTTP Terminator tested 30,000 websites where

A GitHub issue opened by an account with no repository privileges was enough to execute code on the CI runners behind Anthropic's and Google's own coding-agent repositories. On OpenAI's, it was enough to hijack the next agent run. Novee Security ran the attack against each vendor's agent in the configuration that the vendor ships by default, and presented the work at Black Hat USA on August 5.

A new analysis has uncovered that the threat actor tracked as TeamPCP has been active on the cybercrime scene as far back as 2020, indicating the group has been compromising internet-facing infrastructure for years before training their sights on the software supply chain. "The connection is supported by overlapping domains, malware deployment paths, staging techniques, backend infrastructure,

Cisco has rolled out updates to address multiple critical security vulnerabilities impacting Catalyst SD-WAN and IOS XE Software as part of a comprehensive internal security review. The security issues affect Cisco Catalyst SD-WAN Software, regardless of device configuration, and Cisco IOS XE Software when it is running in autonomous or controller mode. "These vulnerabilities were found

Apparently, opening the thing is now enough. A repo can run before the first prompt, a package can hide among hundreds, and a harmless-looking PDF can finish the job. This week runs on cheap leverage: exposed servers, recycled bugs, poisoned agent instructions, remote-access tools dressed as support software, and trusted defaults doing attackers a favor. Nothing here is especially mystical.

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

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

OpenAI said it disrupted a Cambodia-based scam operation that used its generative artificial intelligence (AI) chatbot ChatGPT to facilitate a wide range of investment, romance, gambling, and law enforcement impersonation schemes. To that end, it banned a coordinated network of ChatGPT accounts likely originating from Southeast Asia and operating from the city of Poipet, a region with extensive

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

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

HashiCorp, Veeam, and the Django Software Foundation have patched 11 vulnerabilities across Terraform MCP Server, Veeam Service Provider Console, and Django. The three most serious: An unauthenticated flaw in Veeam's console that hands over a managed agent's credentials, rated 9.5 A cross-tenant flaw in HashiCorp's MCP server that lets one user's Terraform token be reused for later users'

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

A credential-stealing npm worm that first appeared in keyv@6.0.0 spread beyond the Keyv and Cacheable namespaces into hundreds of packages across multiple organizations on August 4, 2026. SafeDep verified 353 poisoned versions across 79 package names in the npm registry. Its monitoring put the wider footprint at 442 versions across 353 names, while Aikido later reported at least 868 packages

The cybersecurity industry has spent decades assuming that offensive capability scales with technical expertise. That assumption is starting to break. Security teams have long estimated risk by ranking attacker sophistication. Nation-state actors sat at one end. Organized criminal groups followed. Inexperienced attackers, dismissed as "script kiddies," sat at the other end, running public

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

This week kept coming back to permission. A model crossed a boundary. A wallet trusted bad randomness. Webmail kept an intruder around. Public systems, package feeds, hotel networks, and login flows all gave away more than intended. Some of it was clever. Most of it was just access left lying around: old bugs, exposed gear, poisoned dependencies, weak defaults, and tooling that moved from

AI is moving incredibly fast, and every security leader is feeling the pressure to keep up. AI platforms like Claude, Codex and Cursor are already helping security teams write detections, investigate alerts, summarize incidents, and automate repetitive work. The conversation has evolved from whether AI belongs in the SOC, to where each type of AI delivers the most value. With so many new AI

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

Palo Alto Networks' Unit 42 says a Chinese-speaking threat actor used DeepSeek through the open-source Hermes Agent framework to launch attacks autonomously. After an initial Telegram instruction, the agent found internet-facing systems and selected public exploits. The researchers recovered no further operator input in the session. The operator, tracked through the aliases knaithe and KnYuan,

Anthropic on Thursday became the latest artificial intelligence (AI) company to reveal that three of its models, including Claude Opus 4.7, Mythos 5, and an unnamed research model, had breached three unnamed organizations during cybersecurity testing without its knowledge. The AI firm said the earliest incidents date back to April 2026, adding it made the discoveries after launching a "

Security experts have been sounding the alarm for years about the risks of using generic TV boxes that promise unlimited content streaming for a one-time fee, warning that they secretly rent the user's Internet connection out to strangers. But a groundbreaking new analysis finds these devices also routinely spoof themselves as mobile phones clicking ads on AI-generated websites as part of sprawling operation that seeks to defraud online merchants and advertising networks.

A lot of security still comes down to trusting the wrong screen. This week, that screen might be a login page, an install guide, a recruiter call, or a familiar service behaving slightly wrong. Behind it: reused credentials, exposed systems, quiet loaders, abused trust, and exploit paths that should have been harder. Some defenses improved. The loose parts still got found first. Anyway,

Hidden instructions in a Word document can make Microsoft 365 Copilot rewrite figures in a report, then copy the same instructions into the finished file. Håkon Måløy disclosed the technique on July 28, 144 days after reporting it to Microsoft. In his proof of concept, the internally generated file triggered the same behavior when it was used in a second Copilot drafting session. Måløy's

Network firewalls are the workhorses of modern cybersecurity. They are trusted to protect the network, blocking malicious traffic and preventing intrusions and breaches. And for decades, network security teams have built controls around a relatively stable model: users connect to applications, applications exchange data, and security tools inspect packets, protocols, and destinations. Firewalls

South Korean authorities and four security firms have disclosed a state-sponsored campaign that compromised trusted domestic websites. The attackers used those sites to exploit locally installed financial-security software and infect targeted visitors with SIGNBT or COPPERHEDGE backdoors. A compromised page could infect a system running a vulnerable AnySign4PC version without a prompt or

The Federal Communications Commission (FCC) added foreign-produced mobile robots and networked power inverters to its Covered List on July 28. The move generally prevents new models from receiving the equipment authorization required for import, marketing, or sale in the US. Previously authorized models can still be sold, and devices people already own are unaffected. Federal purchases and use

Cybersecurity researchers have flagged a maximum-severity security flaw in Ruflo, an open-source agent meta-harness for Anthropic Claude Code and OpenAI Codex, that could result in unauthenticated remote code execution. The vulnerability, tracked as CVE-2026-59726 (CVSS score: 10.0), impacts all versions of the project before version 3.16.3. It has been codenamed RufRoot by Noma Security's

AI is compressing exploit timelines. The real question isn't whether your vulnerability management playbook needs to change, it's which part of it you've been getting wrong all along. The conversation happening in security circles right now goes something like this: Mythos is here. Exploit timelines are collapsing. Does the vulnerability management playbook need to change? The honest answer is

OpenAI on Tuesday revealed the rogue artificial intelligence (AI) agent that escaped its sealed evaluation environment and broke into Hugging Face's production environment also hacked multiple third-party accounts and services as part of the attack. The latest disclosure shows that the security incident, which stemmed from an internal security test, was more extensive in scope than previously

Anthropic says Claude Mythos Preview helped derive an end-to-end key-recovery attack against HAWK-256 and a 200- to 800-fold speedup for an attack on seven-round AES-128. The HAWK attack exploits a previously unused symmetry in the lattice behind the signature scheme. Anthropic's released implementation gives an expected end-to-end runtime of about three hours and 42 minutes on a 96-core server

JFrog has confirmed that OpenAI models exploited a zero-day in self-hosted Artifactory while trying to reach the open internet from a sealed evaluation environment. Artifactory is JFrog's software repository manager. OpenAI says the models then escalated privileges and moved laterally until they reached an internet-connected node. JFrog says it has since developed and released fixes for cloud

STAR Labs has published a Linux kernel exploit that turns an ordinary local user into root on the CentOS Stream 9 build it targeted. The flaw, tracked as CVE-2026-53264 (CVSS score: 7.8), is a use-after-free race in the kernel's network traffic-control subsystem.Researcher Lee Jia Jie said artificial intelligence (AI) helped him find the bug and speed up exploit development. This is local

Microsoft has launched its first cybersecurity-specific model inside MDASH, its multi-model vulnerability identification and remediation harness. The company says MDASH, using MAI-Cyber-1-Flash and GPT-5.4, scored 95.95% on CyberGym. It also claims the configuration costs 50% less than its current best MDASH combination of GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex. Access is limited to approved

NVIDIA and 36 other organizations have formed the Open Secure AI Alliance to develop and share open technologies, techniques, and tools for securing software and artificial intelligence (AI) agents. The 37-member group spans cloud, security, enterprise software, and AI companies, including Microsoft, Cisco, Cloudflare, CrowdStrike, Hugging Face, IBM, Palo Alto Networks, Red Hat, and the Linux

Monday starts with the usual promise that everything is under control. Then the logs wake up. This week, trusted tools crossed lines, old flaws found new work, exposed systems stayed exposed, and attackers kept hiding inside normal-looking services. Nothing looked strange at first. That helped. That is the mood. Here is the full recap. ⚡ Threat of the Week OpenAI Says Its AI Agent Went Rogue