Defences against social engineering. Deepfake technology.

The AI tools that analyse what you really thought about that meeting. Unsafe phones at a US government agency. German MPs told to leave Signal. And fake facial hair defeats age checks.

Breached

Two recent incidents this week underline the need for organisations to protect themselves against social engineering attacks. In one, home security giant ADT exposed data belonging to 5.5 million customers after attackers reportedly used a voice phishing call to manipulate an employee into granting access to its systems. Meanwhile, DigiCert (whose digital certificates help verify that software is legitimate) was breached after attackers repeatedly sent booby-trapped screenshots to support staff.

These attacks succeed because they exploit routine behaviour: answering calls, helping customers, opening attachments and responding quickly. The lures are increasingly convincing, with emails sent via legitimate platforms that defeat automated defences (e.g. PayPal and Facebook below). It’s essential to have tested processes in place, and never to be swayed by an urgent call to action.

• Be sceptical of inbound calls, especially any that claim to be from IT support, security teams or account providers

• Never approve logins or password resets you didn’t initiate

• Treat update prompts carefully: update software from within the app or official site, not via pop-ups

• Verify unusual requests through a second channel you already trust

• Enforce cast-iron processes to authorise financial transactions

Threats

Deepfake: 404 Media has details of a software tool capable of changing someone’s appearance in real time during video calls on platforms like WhatsApp, Zoom, Teams and TikTok. Haotian AI still has some limitations, but it’s worryingly good. Running it requires a high-end computer and video card, and $2,500.

PayPal: Scammers have found another way to use PayPal’s legitimate services to deliver fake emails. The tactic involves changing the subject line so that it warns about a large transaction and includes a phone number to call to dispute the charge. It’s not clear how the attackers managed to do this, but the takeaway is clear: be sceptical about this type of message and, if you’re worried, call a number you have checked is genuine. Malwarebytes

Facebook: A similar tactic has been used to send phishing emails by exploiting a feature in Google’s AppSheet platform. Guard.io says it has resulted in the compromise of some 30,000 Facebook accounts. The messages use a variety of tactics, with some warning an account will be disabled while others offer fake “Blue Badges” or non-existent job opportunities. Attackers are expert at exploiting flaws in well-known platforms to send what appear to be genuine emails: the address line is no guarantee the message itself is authentic.

Conduct: Microsoft says a sophisticated campaign targeted more than 35,000 users across 13,000 organisations in 26 countries by impersonating internal compliance teams and accusing employees of breaching company codes of conduct. The polished emails often arrived via legitimate email services and pushed recipients to open PDFs that led them through fake verification steps before stealing login credentials.

Claude: The interest in AI tools makes it essential to double-check what you’re installing is genuine. Fakes are rife and often promoted through Google adverts. Malwarebytes has details of one of the latest examples which involves a fake installer for Claude Code.

Burnt: Hackers hid malicious software in downloads of DAEMON Tools, a popular app for creating bootable USB drives (i.e. ones that can be used to start a computer independently of its normal operating system). The issue, which had been present for almost a month, has now been fixed. Kaspersky

AI

Deployment: Countries from the Five Eyes intelligence-sharing alliance have published joint guidance on how to deploy agentic AI services safely. As they say, “unlike traditional generative AI, which typically requires human validation, agentic AI systems are designed to operate autonomously, making them a powerful tool.” The guidance is aimed at defence and critical industry sectors, but it’s applicable to any deployment of the technology.

Patches: The UK cybersecurity agency says AI will result in a “forced correction” to address a backlog of technical issues that are the “result of prioritising short-term gains over building resilient products.” The NCSC is urging organisations to prepare now for the arrival of a “patch wave” which will require the deployment of updates more quickly, more often and at scale. The NCSC has for years highlighted the failure of the technology industry to prioritise security. That particular chicken is now coming home to roost.

Emotional: How do you feel about meetings? Love them or loathe them, AI tools can analyse your emotions and report back to your boss. As The Atlantic reports, if you’ve ever been warned that your call “is being monitored for quality-assurance purposes,” chances are the person on the other end is being assessed by what’s called “emotion AI.” Those same tools are now being applied more broadly: one estimate says the global emotion-AI market is expected to triple by 2030, to $9 billion.

Brains: Using AI chatbots for just 10 minutes may significantly impact people’s ability to think and problem-solve, according to a study by researchers at Carnegie Mellon, MIT, Oxford, and UCLA. The issue affected those who depended on the chatbot to provide answers. Using them for hints and clarifications did not have the same negative effect.

Sneaky: Google has been installing a 4GB AI model to all instances of its Chrome web browser without informing users. The model powers features like Chrome’s AI scam detection, autofill and writing assistance. You can disable the feature in Settings | System | On-device AI, which seems to get rid of the file. Why Google hasn’t been more transparent about this defeats us. That Privacy Guy

Going up: The number of replies that repeat Russian propaganda and disinformation has doubled in Anthropic’s Claude chatbot from last year. Newsguard tested the chatbot by asking it questions related to false Russian news stories. 15% of the responses contained Kremlin talking points, compared to last year’s 7%. NewsGuard says that since the start of the US-Israel-Iran war, Claude has also started repeating Iranian propaganda, something it didn’t do before.

Loose: Set AI tools free in the real world and chaos ensues. As a popular British mathematician explains, she built an AI agent and gave it a bank card and a series of tasks. As well as doing a bunch of things it wasn’t asked to do, it also handed over her passwords to a stranger. Meanwhile, in Stockholm, Andon Labs told an AI agent to open a cafe. Which it did, while merrily impersonating real people in order to submit applications for various permits. It also ordered supplies that couldn’t be used and forgot to order ones that were essential. Despite this, the cafe is open, generating $4,700 of sales in its first two weeks, and of course creating publicity for Andon Labs. It’s all quite amusing at the moment. We suspect in a few years this will look like a time of blissful innocence.

Trump

Unsafe: A key agency in the Department of Homeland Security failed to manage employee phones securely, exposing them and their data to malicious software and data theft. Devices did not use recommended security settings and 76% of apps installed on the agency’s mobile devices posed security risks, were prohibited, or allowed prohibited activities. And they weren’t protected for foreign use.

Overreach: Seemingly overlooking the fact that Canada is still a separate nation, the Department of Homeland Security demanded Google hand over personal data, activity, and geolocation information on a Canadian citizen who posted online criticism of the Trump administration after the killing of Renee Good and Alex Pretti in Minneapolis. The man, who hasn’t been named, hasn’t entered the US for more than a decade. WIRED

In brief

Switch: German MPs have been told to move from Signal to Wire for secure communications after the Bundestag President’s Signal account was compromised earlier this year in a phishing attack. The rationale is that Signal’s registration process requires a phone number while Wire just needs an email address. This is all a bit bonkers because Signal now allows users to restrict who can see their phone numbers. The real issue German MPs should worry about is social engineering persuades them to share authentication details.

Travel: The Amadeus travel management platform is buying a leading biometrics and identity services provider. It’s part of a bid to create a “seamless” journey in which faces or fingerprints are used to verify identities from check-in to boarding, and potentially across hotels and border controls as well. Such a system has obvious implications for personal privacy, not least because Amadeus has a far from flawless security record when it comes to protecting passenger data.

GPS: There’s plenty of mischief to be made by spoofing GPS signals, including hijacking trucks. To mitigate the issue, a US laboratory has developed a new technology to detect such attacks.

Ads: AI models can accurately infer sensitive personal traits, including political views, education level, employment status and broader socioeconomic background, simply by analysing the adverts shown to someone online. Australian researchers found that the adverts act like a digital fingerprint, revealing who platforms think you are even if you never explicitly share that information.

Tracker: If a nuclear apocalypse were to be imminent, wealthy private jet owners would probably know about it before the rest of us. That’s the assumption behind the Apocalypse Early Warning System which tracks how many of those planes are in the air and calculates an emergency level from 1 to 5. At the time of publication, it’s sitting at a comforting 1.

Hirsute: Safe to say that UK efforts to impose age checks under the Online Safety Act could be going better. In a survey by Internet Matters, 46% of children said the checks were easy to bypass, including by drawing a moustache on their faces.

Updates

Microsoft: The April 2026 security updates are causing failures in some third-party backup tools. Applications affected include products from Macrium (Reflect), Acronis (Cyber Protect Cloud), UrBackup Server, and NinjaOne Backup. Microsoft says the fix is to update them to the latest version.

Linux: Update any instances of Linux to address a serious vulnerability that affects virtually all releases of the operating system.

cPanel: Attackers are using the vulnerability we reported last week to wipe public websites and post fake ransoms on their homepages. Updates have been released, but too late for many users: the Shadowserver Foundation says more than 44,000 cPanel systems appear to have been compromised in one form or another.

Cisco: Updates to address multiple vulnerabilities across its enterprise products, including five high-severity bugs.

Oracle: So farewell to Oracle’s quarterly updates. The company says use of AI means they need to move to a monthly cycle.

SecureDrop Inbox: Minor update to the new solution for managing sources.

← Back to all editions