← Cybersecurity Today
Cybersecurity Today
Claude Outage Data Leak, Microsoft GitHub Worm, IBM Hack, M Instagram Takeovers, Canada's Bill C-8
00:10:21
Loading summary
Transcript1 lines
- [00:00]
A
Claude goes dark in major Friday outage, and some users say it may have handed them a stranger's conversation. A self spreading worm rips through Microsoft's own code On Friday, a whistleblower says IBM was hacked three times and buried it. Hackers hijack Instagram accounts by simply asking Meta's AI to hand them over. And after years of waiting, Meta Canada finally passes a critical infrastructure cybersecurity law. This is Cybersecurity Today and I'm your host David Shipley. Let's get started. We start with one of the biggest names in AI Anthropic, maker of the Cloud Chatbot. On Friday afternoon, Cloud went down. According to Cyber News, the outage lasted about two hours and hit several of Anthropic's models, including its flagship Opus. At its peak, the Tracksite down detector logged close to a thousand reports in the us an outage alone is sadly routine for Anthropic over the past few months. Here's what made some people nervous during the disruption. A software engineer posted on X that Claude appeared to return another user's response. Someone else's potentially private exchange showing up in their session. That post drew more than a million views. If true, it would be a cross user data leak, exactly the kind of thing that should never happen. Anthropic says it's investigating. The company told Cybernews it has seen no other reports or evidence of customer data leaking. One of its engineers said the cause was an infrastructure outage, not a flaw in how the AI generates answers, and that it would not have touched customer data. Users, however, were split. Some believed cloud really was crossing wires between people. Others figured the chatbot was just producing nonsense. Which AI models can do for a tool this many people and organizations now depend on Reliability is becoming a key part of the story just as they pursue an initial public offering. We've been tracking a threat group called Team PCP since the spring, the Poison Trivi tool, the Cisco Breach, the tainted Axios library, and more. Their self spreading worm is back and it just hit a big target. This time the target was Microsoft. According to the Hacker News, a worm called Miasma hit 73 of Microsoft's own code repositories on GitHub across four of its GitHub accounts, including Azure and Microsoft Docs. GitHub shut down access to the affected repos. Miasma is the latest mutation of a worm Team PCP released back in May, and it's getting nastier. Two things stand out. First, it reinfected a project the group already hit last month Researchers say that's not a coincidence. It means whoever stole the login credentials in May likely never lost them. It's the same wound reopened. Second, the worm now weaponized AI coding assistance. The trap springs when a developer clones an infected project and opens it in an AI coding tool, including Claude Code Cursor and others. The malware runs on its own. Here's why all of this is so hard to stop. The worm doesn't break into npm or GitHub. There's no software flaw here to patch. Instead, it abuses trust. It steals a developer's key, then publishes malicious code that looks completely legitimate. You can patch a vulnerable tool, you can't patch a trust. And open software runs on trust. That's a far deeper hole to climb out of, and we don't have clean, easy answers yet. A former IBM executive says the tech giant was hacked by foreign governments three times and then hit it. The accusation comes from William Barlow. Until 2019, he ran threat intelligence at IBM. He filed a lawsuit back in 2020. It was unsealed last week. Barlow alleges Chinese state hackers broke into IBM's core network between 2013 and 2016. The group is known as APT10. He alleges IBM found out, investigated, and then told no one. Not the public, not the government. In March 2017, intelligence agencies from all five of the Five Eyes countries the United States, the United Kingdom, Canada, Australia and New Zealand warned IBM about the breach. That kicked off an internal probe. The lawsuit alleges that IBM concluded APT10 may have gotten into its network more than 56,000 times, nearly 400 accounts compromised, around 200 systems across 18 countries. IBM is pushing back. It told TechCrunch the complaint is six years old. The Justice Department declined to take it up, and the company is confident it followed the law. None of these allegations have been proven in court. This next story is almost too simple to believe, but it's real. Hackers took over Instagram accounts by just asking Meta's AI chatbot to hand them over. According to TechCrunch, Meta now uses an AI chatbot for customer support. The attacker would tell the bot they owned someone else's account, then asked the bot to link it to an email they controlled. The bot did it. From there, the attacker reset the password and locked the real owner out. No Meta employee was ever involved, no human in the loop. The targets included high profile accounts, among them a senior US Space Force official and reportedly a dormant Obama era White House account, though Meta disputes that one. The big prize was short valuable usernames so called OG handles that get resold like collectibles, Meta said on Monday the problem was fixed by Tuesday. More users reported being hacked. TechCrunch saw hackers in a telegram channel claiming the trick still worked and selling stolen handles. Meta has started emailing victims and forcing password resets. It won't say how many people were hit. Now let me be blunt about the real lesson here. There's a 2025 study out of the University of Pennsylvania's Wharton School called Call Me a Jerk. Researchers ran 28,000 conversations. They found that classic human persuasion tricks flattery claims of authority. Getting the bot to agree to something small first more than doubled an AI chatbot's willingness to break its own rules. Compliance jumped from about a third of the time to requests that it should have refused to nearly three quarters. One tactic hit 100%. You cannot train social engineering out of these chatbots. You cannot stop this with guardrails. They learn from the human language, so they fold to the same pressure people do. It's baked in. So handing a chatbot the power to change accounts or reset passwords is one of the most reckless things an organization can do. Critics have warned about this for years. They keep getting ignored, except by criminals who are putting it to work. Canada finally has a cybersecurity law for critical infrastructure Last Thursday night, the Senate of Canada passed Bill C8. It now goes to the Governor General for royal assent. Full disclosure I've testified about this bill at the House of Commons and the Senate over the last four years in its various forms. I argued passionately for passing it. Here's what C8 does. It sets mandatory cybersecurity rules for four telecommunications, finance, energy and transportation. The government can order telecom companies to drop risky suppliers and pull dangerous gear out of their networks. It forces mandatory cyber incident reporting and Break the rules and you face stiff financial penalties. Some privacy advocates have called this a surveillance bill by stealth. Many cybersecurity experts disagreed, including Aaron Schall, general counsel at the Centre for International Governance Innovation, a Canadian think tank. The argument from the cybersecurity professional side was simple. While the bill isn't perfect, doing nothing while threats keep coming is far worse. Canada was years behind other G7 countries. The privacy Commissioner of Canada flagged broad government powers and what they said were weak rules for sharing data with foreign governments. The University of Toronto's Citizen Lab said the bill still lacked sufficient safeguards. During committee study at the House of Commons, MPs added amendments that would have forced the government to get a judge's sign off before a minister could issue a cybersecurity order. The speaker of the House struck those down, ruling they went beyond what the bill was meant to cover. So those kinds of orders now only face a court after the fact, not before. For cybersecurity advocates, speed was important, particularly when faced with a national cybersecurity critical infrastructure threat. What happens next? Royal Ascent and at least 18 months of writing the actual regulations that will turn the law into reality. That's Cybersecurity Today for Monday, June 8, 2026. Thanks for listening. Thanks for all of your support, for continuously leaving reviews ratings and for sharing the show with others. It truly helps us grow our audience. I'll be back on Wednesday with the latest cybersecurity headlines. Take care and stay safe.