In today’s cybersecurity news…

CISA lacks Mythos access

Sources tell Axios that the Cybersecurity and Infrastructure Security Agency doesn’t have access to Anthropic’s much vaunted Mythos Preview model. Anthropic briefed both CISA and the US Commerce Department on Mythos prior to public testing, and we previously covered a Politico report that the Commerce Department’s Center for AI Standards and Innovation and the NSA have access to testing. This comes as CISA continues to fight for staff and funding. Staffing is down by about a third, 1,000 people, since January 2025, including virtually all top officials, and a proposed budget that would slash funding by 17%. 

(Axios)

Lovable denies data leak

The security researcher known as @weezerOSINT on X disclosed that by creating an account on the vibe-coding platform Lovable, they were able to access other users’ source code, database credentials, chat history, and customer data from other free accounts via the service’s API. A flaw disclosed to HackerOne was declared a duplicate and left open for over a month. In response, Lovable said that part of this stems from its initially unclear documentation, in which publicly disclosed projects intentionally exposed this information. In May 2025, Lovable disabled public settings for enterprise customers and allowed free users to make projects private, switching to default private projects in December 2025. However, it found that in February, a retroactive patch to its API “accidentally re-enabled access to chats on public projects.” Lovable fixed the issue on April 20th. 

(The Register)

YouTube opens up deepfake detection tool

The video hosting giant has been working on this tool for the last three years, and testing it since 2014 with selected politicians and public officials. Now, YouTube is expanding access to anyone at high risk of having their likeness used maliciously. Similar to its Content ID system, users need to opt in to YouTube’s deepfake detection, uploading a likeness. The system will then automatically flag deepfakes. Only those that show “realistic and consequential disparagement” and “content replacement” will be removed. YouTube’s guidelines on parody and satire would still apply. Unlike YouTube’s Content ID system, deepfake detection does not give users the option to split revenue with creators uploading their likeness. 

(The Hollywood Reporter)

Google patches Antigravity IDE

Researchers at Pillar Security discovered a vulnerability in Google’s agentic IDE Antigravity that allows execution of arbitrary binaries. The researchers found that Antigravity did not sufficiently sanitize inputs on its native file-searching tool, which runs before any of the tool’s Strict Mode constraints. Combined with its file-creation abilities, enabled a full prompt-injection attack chain. While the easiest attack path would require compromising a user’s account, an indirect prompt injection could be possible by pulling in maliciously crafted files from an untrusted source. The researchers disclosed the flaw in January, and Google patched it in late February. 

(The Hacker News)

Huge thanks to our sponsor, ThreatLocker

Microsoft warns of Teams abuse

Microsoft warned its seeing an uptick of attacks abusing external Teams collaboration, using remote management software to transfer files to external cloud storage. These abuse commodity tools, making them hard to spot against legitimate applications. A report outlined a typical nine-stage attack chain, starting with an attacker contacting with an external Teams chat posing as IT, trying to get the employee to do a remote support session. From there, the attacker uses PowerShell to assess their access for lateral movement, then drops a small payload for DLL side-loading. Once fully deployed, they use Rclone or similar tools to exfiltrate data. These attacks seem fairly targeted, aimed at extracting valuable information or maximizing lateral movement. 

(Bleeping Computer)

HandyPay used to steal NFC data

Researchers at ESET spotted a new version of the Android NGate malware that trojanizes the legitimate HandyPay app. This allows the attackers to transfer NFC card data to use at another contactless payment location, as well as capture PINs to a C2 server. This campaign seems targeted in Brazil, distributed through websites spoofing state lottery organizations. The campaign has been active since November 2025. Researchers say the malicious HandyPay app shows evidence of being written by an LLM. 

(The Hacker News)

Ransomware negotiator was helping the bad guys

Angelo Martino, a former negotiator at the incident response company DigitaMint, pleaded guilty to charges related to working with ALPHV to deploy ransomware at US companies in 2023. Martino, along with two other DigitalMint negotiators, both of whom pleaded guilty to conspiracy to obstruct commerce by extortion late last year, now face 20 years in prison. While working with five victims, Martino shared information on their negotiating positions and insurance policies with ALPHV operators to maximize profits. The three later served directly as ALPHV affiliates. These attacks resulted in organizations paying over $50 million USD in ransoms. 

(Bleeping Computer)

Iran claims US using backdoors on network equipment

Iranian state media reports that during recent attacks in the US-Israeli war with the country, hardware from Cisco, Juniper, Fortinet, and MikroTik either rebooted or disconnected. The report proposes that the US installed hidden backdoors in the devices’ firmware or bootloaders. Because of Iran’s ongoing internet blackout, this would have had to be activated by a satellite signal at a specific time or through a botnet. As The Register points out, with the ongoing internet outage, there is no way to verify any equipment outage, let alone the cause. We’ve seen the US use cyber operations extensively in other armed conflicts, but apply all the necessary grains of salt to this state media report. 

(The Register)