Apple just released a system called “Reference Image.” It can verify the image is exactly as taken by an iPhone—new models only—without tying it to a specific iPhone or photographer. It can also verify that multiple images came from the same iPhone.

Other industry solutions require a photographer or institution to vouch for an image using their own credentials. We are concerned this puts some photographers, such as those operating in conflict zones, in a difficult position; it should not be necessary to forgo anonymity in order to prove image authenticity. We built Apple Reference Image to avoid using an explicit, public credential for photographers, and to avoid even implicit public association between different photos taken by the same sensor. The final reference image is instead signed by Apple’s signing service, after validation by PCC. That signature is backed by Apple’s strongest technical guarantees.

Our implementation also protects the confidentiality of the image itself, including from Apple. Merely capturing a reference image should never expose the actual pixels to Apple or anyone else. We achieve this through the exceptional privacy properties of PCC—the nodes themselves are architected so that not even Apple can access image data, just as Apple cannot see the information processed for Apple Intelligence in PCC. While the revocation service must maintain a private record of photo GUIDs and associated sensors to allow for revocation, it never has access to the image data, and does not allow for public access to this record. And as final revocation checks occur using on-device lists, a device never reveals to anyone which photo it’s looking at in order to find out whether it’s still valid.

The report makes for good reading; the details are interesting.

Earlier this month, Anthropic published a long report detailing all of the Claude misuses it detected. Daniel Meissler usefully summarized the report into 117 findings.

A few of the highlights:

• AI agents increasingly handled reconnaissance, exploitation, data theft, propaganda production, surveillance workflows, and research while humans selected targets, set goals, and reviewed important outputs.
• The report describes attackers using AI to industrialize credential theft, cloud compromise, phishing, vulnerability research, and the extraction of sensitive data from downstream organizations.
• Influence operations used persistent agent memory, fake news sites, fabricated journalists, synthetic personas, political profiling, and large-scale multilingual content, although high content volume often produced little genuine engagement.
• Surveillance and repression cases included automated dossiers, biometric and communications analysis, transnational targeting, coercive recruitment, and systems that continued operating locally after model access was revoked.
• Biological and weapons cases show dual-use risk: AI supported advanced scientific and military work, but the report generally doesn’t establish completed biological weapons or operational battlefield deployment.

Last week, Anthropic released a long and detailed document describing current misuses of their Claude models. I’m still reading it, but I wanted to flag this:

We identified a cell of threat actors based in northern Yemen running three weapons development programs: a guided rocket that used a commodity phone-class flight computer with final-phase homing guidance; a multi-stage ballistic missile with a stated range goal above 2,000 km; and a multi-variant missile (referred to as the “R2000” set) that included a hypersonic glide vehicle variant.

The actors used Claude Code in place of human software engineers to develop the guidance, navigation, and control (GNC) software that steers and stabilizes a flying vehicle. For example, they used Claude to integrate an open-source autopilot onto a phone-class flight computer, writing the control and position estimation software, tuning the control settings, running a firmware build pipeline, and performing a flight simulation. The actors managed several Claude instances at once, assigning each one a role, much as a lead would delegate work on a small engineering team: the actors tasked one instance with writing the code, another with research, and a third with reviewing the code the first instance produced.

Our safeguards blocked many of their requests, but not all of them. The actors used a variety of tactics to evade our safeguards, including hiding their goals and the products the software was meant for, and they split their work across multiple sessions so no single session revealed their full intent.

These actors carried out a sustained effort to develop guided weapons, including using Claude to design guidance software. We do not have evidence the actors succeeded in fielding an operational device; but they did test-fire a guided rocket. This field test appears to have failed: within hours, the actors returned to Claude to work out why it failed.

Expect more of this. AI systems democratize expertise and capability. Most of the time that’s a good thing, but sometimes it’s not.

The chart is interesting.

On the IPI benchmark, Opus 5 improved over Opus 4.8, reducing the probability of an attacker succeeding within 15 attempts from 5.5% to 2.0%, and from 0.5% to 0.2% on 1 attempt. It also improved on Sonnet 5 (5.9% at k=15) and Mythos 5 (2.6%), making it the most robust model evaluated. Opus 5 also outperformed all non-Claude models on this benchmark. The most robust non-Claude model was Muse Spark at 16.5% within 15 attempts—more than eight times Opus 5’s rate. The most capable GPT 5.6 variant, Sol, was comparable to its predecessor GPT 5.5 (20.0% versus 20.8% within 15 attempts), and was 10 times as likely to be successfully attacked as Claude Opus 5 at 2.0%. The other GPT 5.6 variants are less robust, at 30.4% (Terra) and 43.9% (Luna). A single attempt against GPT 5.6 Sol succeeded 3.1% of the time, higher than the 2.0% an attacker achieved against Opus 5 after fifteen attempts.

We know that preventing prompt injection is impossible in the general case. But we are getting much better at blocking it in specific cases.