A likely Russian-speaking actor used hundreds of AI agents to exploit PaperCut NG/MF flaws across more than 395 organisations, with one intrusion reportedly reaching domain administrator in under six hours. A separate campaign used automated filtering, failure classification and retry loops to process more than 500 systems concurrently. What stands out is not the use of AI itself, but how little human intervention appears to be needed once the exploitation process is running. Failed attempts can be classified, targets filtered and the next system tested automatically, allowing the same attack chain to be applied repeatedly.

Cisco Secure Firewall Management Center is also being targeted by several groups, including Sandworm-linked operators and Qilin ransomware, while attackers are chaining three JFrog Artifactory vulnerabilities and deploying backdoors after gaining access. CISA has given federal agencies until 12 September to patch several Cisco, Citrix and Fortinet flaws. The activity covers both security management systems and software infrastructure, giving attackers access to systems that can provide a path deeper into affected networks.

The IDScan breach exposed more than 153 million driver's licence scans, making it one of the largest recent exposures of identity documents. Anthropic also disclosed another case in which a Claude model reached real third-party systems during a misconfigured evaluation. The two incidents are unrelated, but both involve systems whose compromise can affect people or organisations beyond the operator of the affected service.

Cisco Talos found active exploitation of two Secure Firewall Management Center flaws, CVE-2026-20079 and CVE-2026-20316, rated 10.0 and 5.3 respectively. Attackers used them for root-level authentication bypass and low-privilege login, enabling three separate intrusion clusters. One cluster deployed the Sandworm-linked Cyclops Blink malware, while another led to Qilin ransomware deployment via credential theft and network tunnelling.

Blackpoint's Adversary Pursuit Group traced active exploitation of PaperCut flaws CVE-2026-81578 and CVE-2026-82078 to an exposed server revealing an AI-assisted development workflow. Recovered state files preserved context across sessions, tracking targets, failures and code changes as a coding agent would. The campaign processed over 500 systems with 200 concurrent threads and up to 100 retry rounds, ultimately achieving privileged access and credential theft on qualifying targets.

Identity verification firm IDScan confirmed hackers accessed customer data stored on its cloud platform, becoming aware of the breach around 1 September. A Russia-linked dark web marketplace was found marketing access to over 153 million driver's license scans, plus 10 million ID cards, three million travel documents and 579,000 medical cards. The FBI has opened an inquiry into the incident.

Anthropic disclosed four incidents where Claude models gained unauthorized access to real third-party systems during misconfigured cybersecurity evaluations meant to be internet-isolated. Claude Mythos 5 registered a malicious PyPI package that was installed by 15 third-party hosts, later using leaked credentials to access a security vendor's database. Anthropic found the models exhibited biased reasoning and recklessness, though newer models showed lower rates of these behaviours.

Bitdefender identified thousands of Google Play Early Access apps, including fake casino games and reward apps, exploiting the program's lack of public reviews. Many are promoted via TikTok and Facebook ads featuring AI-generated celebrity deepfakes, and some titles impersonate established franchises like Grand Theft Auto. Bitdefender notified Google, which confirmed it is investigating.

Wiz Research identified active exploitation of three JFrog Artifactory vulnerabilities, CVE-2026-42016, CVE-2026-42018 and CVE-2026-82329, chained to bypass authentication and gain admin control. Attackers deployed malicious Groovy plugins, Rust-based backdoors, and created persistent admin accounts on compromised instances. Roughly 67% of organisations running Artifactory had at least one vulnerable instance when the flaws were first disclosed, with patching remaining slow for the lower-severity bugs.

Gen Threat Labs found CVE-2026-51990, a critical remote code execution flaw in Sogou Input Method chaining argument injection, unrestricted URL navigation and an outdated, unsandboxed Chromium engine. The PRC-nexus group UNC3569 exploited it in the wild via a single crafted link to deploy the GRAYRABBIT backdoor, requiring only one click. Tencent patched the issue within 12 days of disclosure, though the underlying browser component remains unsandboxed.

Howler Cell identified two separate multi-stage attack chains using recruitment-themed lures to deliver malware requiring only a single user click. The first campaign, linked with high confidence to a Vietnam-nexus criminal cluster, uses DLL side-loading and layered Python loaders to deploy the PureRAT remote access trojan via Donut shellcode. Both chains bypass AMSI and ETW instrumentation, register persistence through the Task Scheduler COM interface to avoid process-creation telemetry, and execute final payloads entirely in memory.