The email your assistant read was not the one you were sent
Researchers hid instructions inside an ordinary email using invisible text and hijacked an AI summariser ten times out of ten, moving a payment deadline and inflating the amount. Elsewhere, eighty British performers asked for legal ownership of their own voices.
- 01
Invisible text in an email rewrote what the AI summary said
Forcepoint published a lab test on 25 August in which instructions were placed inside an email in white, zero-height type, invisible in Outlook but sitting in the HTML that gets passed to the model. An ordinary summariser with no guardrails then followed them: in all ten runs the payment deadline moved from 21 August to 3 September, the outstanding amount came back more than five times the real figure, and a named contact vanished from the summary. Roughly half the characters in that email were never shown to a human. Nothing in the output indicated the summary had been tampered with.
Forcepoint X-Labs - 02
Eighty British performers want your voice made a statutory right
Around eighty performers, among them Hugh Bonneville, Nicola Coughlan, Matt Lucas, Luke Evans and Siobhán McSweeney, signed an open letter to the Prime Minister on 28 August backing a campaign called Save Our Voices Now. It asks for a person's voice to carry legal protection in the way a name or an image already can, on the grounds that three seconds of recorded audio is now enough to clone one. An open petition on the parliament site making the same case, Strengthen copyright laws on our voices, runs until 6 January 2027 and has passed 2,500 signatures. UK law currently has no statute covering a synthesised copy of your voice, which is worth knowing if yours is on a company video, a voicemail greeting or a site walkthrough.
UK Government and Parliament Petitions - 03
Meta glasses now stop filming if the recording light is covered
Alex Himel, who runs Meta's wearables division, posted on Threads on 27 August that a fix is rolling out for people who start a recording and then cover the capture LED with a finger or tape. The glasses already refused to start filming with the light blocked; now the camera stops mid-recording if it gets covered. This is the second such patch in under two months, after an update that disabled the camera when the LED had been physically damaged. Anyone running a site or an office where visitors turn up in smart glasses is relying on that small light doing its job.
Engadget - 04
More than a hundred firms say AI attacks are about to get much more common
A joint letter published on 27 August and signed by over a hundred organisations, including OpenAI, Anthropic, Google, Microsoft, AWS, Mastercard, Visa, CrowdStrike and Cisco, argues there is a narrow window to strengthen defences before AI-enabled attacks become routine. It asks governments to fund protection for critical infrastructure that is under-staffed and under-budgeted, and asks the AI companies themselves to hand defenders controlled model access, funding and hands-on help. The phrasing worth noting is that current security is not expected to hold. Nothing in it obliges anyone to do anything.
OpenAI - 05
Two men charged over the supply chain worm that hit a thousand organisations
The AFP, working with Western Australia Police and the FBI, charged two men aged 21 and 23 on 26 August with a combined fourteen offences over their alleged role in TeamPCP, the group behind the self-spreading Shai-Hulud worm. The method was to poison widely used open-source packages so that anyone installing them handed over credentials, which were then used to poison more packages. Reporting puts the reach at more than a thousand organisations, with hundreds of thousands of credentials taken. None of those firms did anything wrong beyond updating a dependency.
Australian Federal Police - 06
A judge threw out the Pentagon's supply chain risk label on Anthropic
US District Judge Rita Lin vacated the Defense Department's designation of Anthropic as a national security supply chain risk, in a ruling reported on 28 August, finding it retaliation for the company publicly refusing to strip safeguards against mass surveillance and autonomous weapons. Lin wrote that invoking national security is not a blank cheque to punish critics. A separate appeal over a related designation is still running, and the government is expected to challenge this one. The practical read for a buyer is that a vendor can be blacklisted and unblacklisted by forces that have nothing to do with the product.
NOTUS
The Forcepoint test is the one to sit with, because it breaks an assumption almost everybody is now making without noticing. You have an assistant summarising your inbox. You read the summary instead of the email, which is the entire point of having it. Forcepoint put instructions into an email in white type at zero height, so Outlook shows nothing and the model reads everything, and ten times out of ten the summary came back saying what the hidden text told it to say. A payment deadline pushed nearly a fortnight later. An amount more than five times what the sender actually asked for. A name quietly dropped. No warning, no odd formatting, no clue on the screen that half the email was never meant for you.
What makes this different from ordinary phishing is that nothing asks you to click. The bait is aimed at the software, and the software passes the result to you in the calm, tidy voice you have learned to trust. From direct experience in construction, the invoices that cause real damage are never the obviously wrong ones: they are the ones that look exactly like the last forty, so nobody stops. An assistant that quietly restates figures is a machine for producing that feeling on demand. The fix is dull and it is the same fix as ever: money and dates get read off the original document, not off a summary, and anything that moves a payment gets confirmed on a channel the email did not choose for you. Use the summary to decide what deserves your attention. Never to decide what you owe.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.