Some attacks take real skill. Others just need someone to leave a door open. This week has both. Exposed passwords, old bugs still getting exploited, and trusted software doing things it shouldn't. There's a strange mix of clever tricks and careless mistakes.

Hackers are borrowing trust from familiar apps, websites, and AI tools. Some are using AI to speed things up. Others are leaving their own attack files exposed. Not exactly careful work. Different stories, familiar weak spots. Here's what showed up this week. The threats change every week.

, and we’ll alert you when each new ThreatsDay Bulletin is out. Hudson Rock's investigation into the FortiBleed credential harvesting campaign has uncovered 975 newly identified domains and 40,573 exposed secrets, including 37,076 local-user records and 1,576 Active Directory LDAP bind records.

The data comes from newly obtained 1,703 FortiGate source files associated with 1,689 host identifiers. "The configuration material dates to around February 2026, earlier than the likely May/June material examined in the first investigation," Hudson Rock said . The latest discovery has been codenamed FortiBleed 2. 0.

Bank of America is reminding customers not to share banking credentials with AI agents.

In an email blast to customers titled "Protecting your finances in the AI World," the financial institution urged customers to never share their username and password with another person or an outside service, "including an AI-enabled tool or agent that offers to act on your behalf."

Google has announced that an upcoming Android release will no longer be able to directly write to contacts with the WRITE_CONTACTS permission. "Instead, you can use the standard contact intents (ACTION_INSERT, ACTION_EDIT, and ACTION_INSERT_OR_EDIT)," Google said .

"Currently, if your app needs to save even one contact, you might be requesting WRITE_CONTACTS. But that permission grants broad access, allowing an app to make widespread changes to contacts across the device in the background, even without user interaction. Most apps don’t need such broad permission." A scan of 7.

25 million public repositories on Gitee, China's largest indigenous Git hosting platform, has uncovered 26,400 unique verified-live credentials, including an active organization-admin credential for a major enterprise AI platform and thousands of live cloud provider access keys, independent security researcher Ben Zimmermann said .

Fortinet has released a "coordinated set of security updates for FortiOS, FortiManager, and FortiAnalyzer" to address critical and high-severity flaws that were flagged using an AI-assisted vulnerability discovery approach. Details of the issues are being kept under wraps until November 10, 2026.

With Fortinet devices becoming a favorite target for attackers, it's essential that users move quickly to apply the fixes.

Details have emerged about a now-patched flaw in Microsoft Authenticator ( CVE-2026-80097 aka AuthMask) that could be exploited by installing and running a malicious application on an affected Android device and convincing the user to complete an authentication flow in the app by claiming to be a trusted app and getting the corresponding tokens.

The flaw could be exploited to take over work accounts without requiring additional permissions. "Successful exploitation could allow the attacker to obtain authentication tokens and access resources as the affected user," Microsoft said.

According to Enclave , "an unprivileged Android app with zero permissions could impersonate any installed Microsoft first-party app and have the Microsoft Authenticator broker hand it that app's tokens. The setup is a device with Authenticator installed and one work account signed into it."

Attackers are disguising themselves as AI agents to steal credentials. Cequence said it has tracked a single actor running AI agent impersonation at scale by posing as about 32 different AI bots, including ChatGPT, Claude, and Perplexity, while searching for passwords, API keys, and cloud credentials.

"The malicious scanner runs from a major cloud provider and rotates through about 32 AI bot identities, spanning crawlers and assistants from OpenAI, Anthropic, Meta, Perplexity, and others," Will Glazier said . "It ignores web pages and requests files that commonly store secrets, including .

env configuration files, cloud credential files, private SSH keys, and server environment data, working from a list of roughly 90 file paths. Each address sends a short burst of 50 to 190 requests in about eight seconds and then moves on.

The scanner works through targets one at a time, around the clock, and first sends random nonsense paths to find websites that answer every request, a common scanning technique.

It hit organizations in financial services, retail, telecommunications, energy, and software, and stayed active from at least late September through late October." The U. S. Embassy in Fiji has revealed that hackers linked to China broke into a regional organization based in Suva between July and August 2026.

"Hackers with links to China maintained access to the organization's network over several days," it said. The embassy did not name the threat actor. Nor did it disclose what data the attackers may have accessed.

Attackers have been breaking into enterprise networks via camera management software, according to new research from security firm Huntress, which observed multiple incidents in September and October 2026 and traced them back to a compromised application.

Source code artifacts point to a Korean company selling analog and IP cameras and digital and network video recorders. The activity abuses the legitimate "WpcMon. exe" binary to sideload DLLs that establish outbound connections to a remote server.

"This activity demonstrates how attackers can blend malicious code with trusted Windows components to evade detection," Huntress said .

In a major operational security blunder, an unknown threat actor stole 68 GB from a Philippine government agency, maintained a detailed log of the transfer, then left that log open on the internet at 13. 163. 205[.] 35:80. "The copy finished in under 24 minutes on September 6, 2026," said .

"The transfer log records 27 SQL Server backups whose filenames correspond to the case-management systems of a Philippine overseas-employment and migrant-worker administration. Beside the log sat the operator's toolkit, a rebranded GodPotato build, a PowerShell stager, two payloads and a full Impacket environment."

Canadian AI company Cohere has claimed that it caught Alibaba's Qwen3.

8-Flash-Next model unsuccessfully attempting to send data to China from its evaluation sandbox as part of a prompt instructing the LLM to detect malicious file download and execution by monitoring for curl processes downloading bash scripts to /tmp followed by bash execution.

"Asked to write a defensive detection rule, Qwen3. 8-Flash-Next ignored instructions and inventoried the machine it was running on and then tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing," Cohere said . "We cannot answer whether this is deliberate or a leftover glitch from training."

Cohere has also released a dataset containing the model trace and corresponding raw token IDs in connection with the behavior.

"The exact output is not guaranteed all the time, and many trials may be required to reproduce the behavior," Cohere said, adding "there are similarly unexpected URLs that the model also tends to output, most containing aliyunc."

However, the findings have also led to social media chatter that the company is resorting to geopolitical fearmongering and that they are an attempt to clamp down on open-weights models. The U. S.

National Institute of Standards and Technology (NIST) has released guidelines that detail the timelines for transitioning away from quantum-vulnerable algorithms to quantum-resistant standards. "The report describes different use cases an