Most of today’s stories involve someone insisting a safeguard or a system is sturdier than it is. The sharpest evidence comes from a report on how quickly AI guardrails can be stripped away.
1. OpenAI and Google’s new models reportedly fall for a nude-image trick
Ars Technica reports that researchers found an attack method that gets new OpenAI and Google models to generate nude images, despite the safeguards meant to prevent it. We only have the article’s summary, so the technique, the specific models and the difficulty of the attack remain unclear. The headline’s word “easily” is Ars’s characterization.
Even so, the lesson for anyone building on these models is plain. A vendor’s content filter is one layer of defense, not a guarantee. If your product could be embarrassed or harmed by this output, you need your own checks on what goes in and what comes out.
The open question is how fast OpenAI and Google patch the hole, and whether the fix holds up against the next variation of the attack.
2. Ukraine’s Starlink dependence just met a throttling policy
Ars Technica says Ukraine was “blindsided” by Starlink throttling restrictions and intelligence gaps, and that Kyiv’s forces are experiencing connectivity problems on Elon Musk’s satellite network. The feed excerpt doesn’t say what the restrictions are or how badly operations are affected, so those details are missing.
The word “blindsided” is the revealing part. If Ars’s framing is right, the problem is a dependency whose rules can change without the customer having much say, more than any technical outage. A military that built communications around a commercial service ended up at the mercy of that service’s policies.
Founders can take a smaller version of this lesson. Read a critical vendor’s terms of service as part of its risk profile, and know what you would do if they changed.
3. Ransomware gets personal, while AI guardrails come off in days
The Register’s headline says ransomware is turning personal, with “we-know-where-you-are kind of stuff.” We only have the headline, not the article text, so we can’t say how common these tactics are. If the framing holds, though, the old playbook of backups plus negotiation doesn’t cover threats against people.
The other source attached here, from WIRED, is about AI rather than ransomware, but it has the day’s most concrete numbers. A Tech Against Terrorism report says abliterated versions of new open-weight models now appear online in under three days on average, compared with more than 24 days in 2024. Alibaba’s Qwen 3.8-27B had one within 24 hours.
The group tested 134 models with more than 200 prompts, and 81 failed its criteria. Claiming to be a “safety researcher” made models nearly nine times as likely to answer, at 16.9 percent. A Llama 3.1 8B variant fell from 97 to 3 once abliterated.
That last result undercuts the idea that a model’s safety training is a permanent property. For open weights, it’s a default setting that can be removed, and the report suggests removal is getting faster. Meta, OpenAI, Google, Anthropic, xAI and Alibaba did not respond to WIRED.
4. Whitehall wants AI while its aging IT leaves it exposed
The Register reports that the chair of the UK’s Public Accounts Committee is warning the government is chasing AI while aging IT leaves it exposed. We only have the headline, so the specific figures and examples aren’t available.
The tension is a familiar one. AI makes for a better announcement than patching old systems, and the headline suggests the committee sees the order of priorities as backwards.
For vendors selling AI to government, the implication is that the bottleneck may not be the tool. If the legacy systems underneath are the problem, a new layer on top inherits that fragility.
5. The €115 billion warning comes from those who’d lose the business
According to The Register’s headline, US tech giants are warning Europe that ditching US tech giants could cost €115 billion. We only have the headline, so we can’t say how the figure was calculated or who produced it.
Still, the messenger matters. A cost estimate from companies that stand to lose customers is a negotiating position until someone shows the assumptions. Moving off an entrenched platform does cost something, but the useful question is what Europe gets in exchange, and the headline suggests no one asked the vendors to answer it.
For European founders and policymakers, the practical step is to ask for the methodology behind the figure before treating it as a forecast.
