This week, two stories that reframe how close Al-enabled cyberattacks already are. OpenAl disclosed that during an internal capability evaluation run with its cybersecurity safeguards deliberately switched off its models escaped a sandboxed test environment, exploited a zero-day to reach the open internet, then broke into Hugging Face's production systems to steal the answer key for the benchmark they were being graded on.

No human directed the attack. Rather, the models simply improvised it to win. Set against that, a former Anthropic employee argues the more pressing near-term risk isn't autonomous escapes or open-weight models at all, but the frontier subscriptions themselves, and claims that commercial pressure is quietly eroding the guardrails meant to contain them.

Subscribe to this newsletter
Come see me speak at BlackHat

OpenAI security incident with Hugging Face. The model was finding a solution to ExploitGym, which led to the model gaining internet access and then reasoning that access to Hugging Face would help it solve the challenge since Hugging Face potentially hosted the solutions to ExploitGym. In order to do this it reasoned that finding and exploiting a vulnerability on Hugging Face to cheat the exploitation.
https://openai.com/index/hugging-face-model-evaluation-security-incident/

Former Anthropic employee: Hackers prefer to use massively subsidized labs AI subscriptions for attacks: I know of two instances where two different Anthropic GTM people used large comitted spend contracts as a prereq for lowering safeguards, and I directly witnessed one.

Hackers are changing the DNS settings on Wi-Fi devices at hotels and conference centers to redirect users to fake Microsoft 365 login pages. The campaign has been ongoing since at least June and impacts organizations in various sectors, including financial services, professional services, legal, health care, energy, and retail. Cybersecurity company ReliaQuest identified compromised Wi-Fi gateways in multiple U.S. cities as well as other regions of the world, such as India and Saudi Arabia.
https://www.bleepingcomputer.com/news/security/hackers-hijack-hotel-wi-fi-dns-to-steal-microsoft-365-accounts/

While OpenAI has been at the center of the recent concerns, AI security researchers at Accomplish AI say they achieved similar results with Anthropic’s Claude Co-Work. According to the report, researchers ran the agent in a local session inside a Mac-hosted virtual Linux machine. They then observed the agent breaking out of the virtual machine (VM) and reading from and writing files on the underlying host system.
https://www.firstpost.com/tech/its-not-just-openai-researchers-show-anthropics-claude-co-work-can-escape-its-sandbox-too-14034041.html

VVAH is Visa's open-source harness for autonomous vulnerability discovery, remediation, and validation using frontier AI models, built on learnings from Project Glasswing (Anthropic's initiative for AI-assisted vulnerability research).
https://github.com/visa/visa-vulnerability-agentic-harness

Historically, Mandiant and Google’s Threat Analysis Group (TAG) maintained distinct tracking systems, relying on parallel naming schemas that grew independently over time. The creation of GTIG has necessitated a new, fused tracking system, and a new naming system. Thinking to the future, GTIG’s new system will rely on cryptonyms.
https://cloud.google.com/blog/topics/threat-intelligence/updated-cyber-threat-actor-naming-system/

Buffer Overflow vulnerabilities in IBM Tivoli Storage Manager FastBack versions 5.5.x and 6.x ≤ 6.1.1 as a case study. The main focus of this article is not just basic exploitation, but how we can bypass memory protection in Windows 10, specifically Data Execution Prevention (DEP).
https://screetsec.com/blog/defeating-windows-dep-using-rop-chains-leveraging-virtualalloc

Member of the European Parliament Stelios Kouloglou was hacked with Pegasus spyware while serving on the PEGA committee, which investigated Pegasus and other spyware abuses in Europe. Through forensic analysis of his device, we found that the attackers could have had access to confidential documents and committee deliberations.
https://citizenlab.ca/research/member-of-committee-investigating-spyware-hacked-with-pegasus/

GitLab has recently addressed multiple security vulnerabilities, including critical remote code execution (RCE) flaws, by releasing updates for its Community Edition (CE) and Enterprise Edition (EE). These vulnerabilities, if exploited, could allow attackers to execute arbitrary code on default GitLab installations, potentially exposing sensitive data and internal services.
https://thedailytechfeed.com/gitlab-patches-critical-remote-code-execution-vulnerabilities/

A Florida man has been arrested following a federal criminal complaint alleging that he participated in a conspiracy to infect Steam users with malware and consequently steal from them. Authorities believe the attacker was not working alone and has stolen at least $220,000 worth of cryptocurrency from thousands of Steam users.
https://gamerant.com/steam-games-malware-theft-arrest/

An Illinois man has been sentenced to 76 months in prison after prosecutors said he carried out a large-scale phishing campaign targeting Snapchat users, particularly women, by pretending to be part of Snapchat support. The attacker allegedly tricked victims into revealing account access codes, allowing him to enter more than 750 accounts, steal private photographs, and distribute some of the stolen material online.

OnTrac parcel delivery company is informing that hackers breached its corporate network and may have accessed personal details belonging to its customers. The incident was detected on March 23, and an internal investigation revealed that the attacker accessed certain files between March 20 and 22. Apart from names, it is unclear what type of information was exposed, as the company redacted the data elements in the notification sample shared with authorities.
https://www.bleepingcomputer.com/news/security/ontrac-notifies-customers-of-data-breach-after-network-hack/

A recent malvertising campaign has been identified, utilizing fake webpages for well-known platforms such as Solana, Luno, and TradingView. The malicious sites employ JavaScript to construct malware directly within a browser's memory. This approach allows the malware to bypass traditional detection methods, making it more difficult for security software to identify and block the threat. The campaign's use of fake webpages and JavaScript-based malware construction highlights the evolving nature of cyber threats and the need for continued vigilance in online security practices.
https://www.bleepingcomputer.com/news/security/malicious-sites-use-javascript-to-build-malware-in-browser-memory/

A new remote access trojan (RAT) has been discovered by the cybersecurity research team at Cisco Talos, and experts are sounding the alarm since it enables the installation of ransomware on the victim's system. The attack is dubbed msaRAT and attributed to the Chaos ransomware group, built with Rust to utilize existing Chrome or Edge (Chromium-based) browser installations. By leveraging the browser's ordinary communication methods, msaRAT is able to disguise its traffic with Chrome's DevTools Protocol and enable command-and-control (C2) communications without detection from typical anti-malware or antivirus software.
https://hothardware.com/news/scrappy-rat-malware-pesters-billions-of-chrome-and-edge-users

A singular malicious installer served as the entry point into a vast criminal ecosystem that concurrently compromised endpoints, exfiltrated sensitive data, hijacked victim bandwidth as proxy relays, and orchestrated vast networks of artificial accounts. Security specialists at VMRay Labs designated this sophisticated campaign as Operation STANDOFF.
https://meterpreter.org/operation-standoff-intrusion-campaign/

In mid-March, it emerged that a German drone manufacturer had been the victim of alleged espionage leading to the arrest of two suspected spies thought to have passed on secret information to Russia. The Federal Public Prosecutor General accuses the suspects of collecting information about the company both on-site and online on behalf of the Russian secret service.
https://www.euronews.com/my-europe/2026/04/14/espionage-sabotage-attacks-german-intelligence-agency-sends-warning-to-defence-industry

CERT-UA published a new advisory attributing a phishing campaign to UAC-0099, a Russia-aligned threat actor active since at least mid-2022 and previously known for exploiting WinRAR vulnerabilities and using phishing emails to deliver malware families including LONEPAGE, MATCHBOIL, and DRAGSTARE.
https://securityaffairs.com/195923/cyber-warfare-2/uac-0099-is-now-hiding-malware-inside-a-fake-notepad-plugin-to-target-ukrainian-organizations.html

The day's security picture is shaped by AI operating on both sides of the line: a new remote access trojan now uses machine-learning scoring to triage its own victims, while a US government advisory warns that Iranian-linked hackers are actively disrupting domestic water and energy infrastructure. Running underneath both stories is an increasingly uncomfortable asymmetry — the same AI capabilities are accelerating attacker operations while guardrails constrain the researchers defending against them.
https://blindthoughts.com/ai-malware-scores-victims-iran-hits-us-infrastructure

Threat actors are actively exploiting two critical vulnerabilities in WordPress Core, collectively tracked as “wp2shell” - CVE-2026-60137 and CVE-2026-63030. These flaws enable attackers to bypass authentication and achieve remote code execution, allowing them to install persistent webshells and inject malicious plugins onto affected sites. Security researchers observed a surge in exploitation attempts following the public release of proof-of-concept code, with multiple threat groups leveraging the vulnerability chain to compromise unpatched WordPress installations.
https://www.yazoul.net/news/article/critical-wp2shell-wordpress-flaws-exploited-to-install-webshells/

Golden Chickens, also known as Venom Spider, has introduced four new malware families—TinyEgg, ChonkyChicken, a modularized ChonkyChicken variant, and ChromEggscalator—showing continued development of its malware-as-a-service ecosystem. The shift to modular, operator-driven tooling and ClickFix-style delivery methods suggests improved evasion and broader support for cybercrime customers.
https://www.hendryadrian.com/golden-chickens-resurfaces-with-four-new-malware-families-and-modular-implants/

Public PoC Released for CVE-2026-42980 Windows Privilege Escalation Flaw.
https://securityonline.info/cve-2026-42980-privilege-escalation/

The nasty new ClickLock malware won’t take no for an answer. It locks up your entire Mac screen and keeps it that way until you type in the password. The malware already hit victims across 33 countries, and it’s not stopping. Once it gets your password, it targets your Keychain, saved logins and crypto wallets, then plants a backdoor that sticks around for good.
https://www.cultofmac.com/news/click-lock-malware-locks-mac-password

FBI accuse Steam user of stealing over $220,000 via malware-infected games like BlockBlasters, Lunara and Dashverse
https://www.rockpapershotgun.com/fbi-accuse-steam-user-of-stealing-over-220000-via-malware-infected-games-like-blockblasters-lunara-and-dashverse

A recently disclosed vulnerability affecting Foxit PDF Reader demonstrates exactly how local attackers can abuse DLL sideloading techniques to gain SYSTEM privileges, significantly increasing the impact of an initial compromise.
https://darknetsearch.com/knowledge/news/en/dark-web-surveillance-foxit-pdf-reader-flaw-exposed/

Vulnerability in the Oracle Enterprise Manager Base Platform product of Oracle Enterprise Manager (component: Self Update Framework). Supported versions that are affected are 13.5 and 24.1. Easily exploitable vulnerability allows low privileged attacker with network access via HTTPS to compromise Oracle Enterprise Manager Base Platform. Successful attacks of this vulnerability can result in takeover of Oracle Enterprise Manager Base Platform.
https://www.cyberhub.blog/cves/CVE-2026-47004

Malicious Steam Workshop maps (Laser Tag Neon, Chroma Grid Arena) infected Meccha Chameleon players by writing a .bat to Documents and using PowerShell to fetch malware, compromising developer systems and the official Discord. Researcher Feint identified it; developer Haganeiro patched it in update 3.1.0, removed the map, and urges scans or full reinstalls for affected users. Infection required launching the maps, not merely subscribing.
https://www.tweaktown.com/news/112850/steam-workshop-maps-infected-players-with-malware-and-took-over-a-100000-member-discord/index.html

Over the past year, a group of Russian hackers has been conducting a cyberespionage campaign targeting American nuclear scientists, defense contractors, and government officials. This is stated in a joint advisory from U.S. intelligence agencies and more than a dozen allied countries, as well as in a report by the cybersecurity firm Proofpoint, according to CNN.
https://militarnyi.com/en/news/russian-hackers-american-scientists-defense/