October 2026 security update: AI techniques go mainstream as edge devices stay exposed

Today’s threat landscape 

AI techniques go mainstream: Cybercriminals steal AI access & hide phishing in invisible text

September showed how quickly cybercriminals adapt, and AI now sits at the center of it. Anthropic's September threat intelligence report shows AI services themselves becoming part of the supply chain. A Russian-speaking, financially motivated cybercriminal injected malicious instructions into an AI vendor's automated testing environment to extract API keys, the credentials that let software use a paid AI service. The cybercriminal then automatically switched the attack workload onto the victims' keys, turning stolen credentials into free computing power and a disguise. The attack continued against more than 30 other AI companies over four days. Across its cases, Anthropic also found that capable groups can modify their malware faster than defenders deploy detection signatures.

Email security is seeing a similar crossover. Microsoft Threat Intelligence found that ASCII smuggling, which hides content in invisible Unicode "tag" characters, has moved into mass phishing. Until now, it mainly served to feed AI assistants hidden instructions. Cybercriminals inserted these invisible characters inside finance keywords such as "funding" and "loan". The characters split the words, so keyword-list filters no longer recognized them while recipients still saw ordinary text. As BleepingComputer reported, the campaign targeted business finance decision-makers with fake loan and credit offers and peaked at more than 2.3 million messages a day.

Defenses built around a single checkpoint, such as a keyword filter or a detection signature, no longer hold on their own. Microsoft's mitigation guidance sums up the email lesson as "normalize before you match": strip invisible tag characters before applying detection rules, and treat their presence as a high-confidence warning sign because they rarely appear in legitimate mail. Microsoft also advises applying the same cleanup to content passed to AI systems.

Exploited vulnerability trends 

Edge devices under attack: Nearly half of September's exploited vulnerabilities sit in VPNs, firewalls & gateways

Devices that guard the edge of corporate networks are where cybercriminals are getting in. CISA added 43 vulnerabilities to its Known Exploited Vulnerabilities catalog in September, nineteen of which sit in remote-access (VPN) gateways, firewalls, email gateways and routers.

Every entry describes an attack happening now, because CISA adds a vulnerability only when it has reliable evidence that cybercriminals are using it. Most of September's edge vulnerabilities also work without login credentials, so cybercriminals only need to reach the device over the internet. These devices make attractive targets because they face the internet by design, everything behind them trusts them, and they rarely run endpoint security software. Once cybercriminals control a gateway, they can capture usernames and passwords as employees connect and spread through the network. They can also stay hidden after the vendor's patch is installed.

Two September entries show what this looks like in practice. In Citrix NetScaler (CVE-2026-88771), cybercriminals can run commands on the remote-access gateway without logging in. In Cisco Secure Email Gateway (CVE-2026-76461), a specially crafted email can trigger SQL injection in the filter that screens incoming mail. For both, CISA asks organizations to patch and check the devices for signs of compromise.

ENISA's EU Vulnerability Database (EUVD) marked 48 vulnerabilities as exploited in September: all 43 from CISA plus five reported by European security teams. These include a MikroTik RouterOS SSH login bypass (CVE-2026-67276) flagged by CERT-PL. It lets attackers log in to a router without the private key and, chained with a second vulnerability, take full control of it.

Organizations in Europe that follow only the US list risk missing the threats closest to home.

You can read more about the EUVD in our blog post here.

Industry news 

Europe sets reporting rules: Cyber Resilience Act obligations start as attackers keep exploiting trust

On September 22, ENISA published its 2026 Threat Landscape, warning that reliance on interconnected digital services expands the attack surface. Supply chain and third-party compromises, the agency found, typically lead to large-scale incidents. ENISA counted more than 48,000 new vulnerabilities published in 2025, a 22% increase on the previous year, and found that 60% of unauthorized access incidents began with a vulnerability. It also found that 73% of targeted organizations were essential or important entities under the NIS2 Directive.

The EU is also putting legal weight behind product security. On September 11, the first obligations of the Cyber Resilience Act (CRA) took effect. Manufacturers of products with digital elements must now report actively exploited vulnerabilities and serious incidents through ENISA's central reporting platform. As Sweden's National Cybersecurity Center explains, they owe an early warning within 24 hours and a fuller report within 72 hours.

RISE warned, as Computer Sweden reported, that tens of thousands of Swedish companies are affected, many without realizing it. Affected companies range from industrial firms and connected-product makers to energy and real estate companies and their subcontractors. Non-compliance risks fines, market restrictions and exclusion from supply chains, and the regulation applies in full from December 11, 2027.

Meanwhile, social engineering keeps working against large institutions. Revolut confirmed that cybercriminals used a government email domain to submit fake emergency data requests, as Computer Sweden reported. They obtained passport copies, addresses and transaction histories of high-profile cryptocurrency customers. Whether through a third-party compromise or a fake official request, attackers keep reaching organizations through the parties they trust.