OpenAI has updated its defenses against adaptive extraction attacks on GPT-6.2 Astra and GPT-6.2 Sol, reducing the success rate of attempts to pull protected internal reasoning from 54.0 percent to 0.2 percent on prior requests, according to The Decoder. Attacks targeting active requests fell from 22.9 percent to 0.9 percent. The company is also rolling selected improvements back to GPT-6 Astra. The results come from internal automated tests and self-reported metrics aggregated across six attacker models.
These updates cover distinct security challenges across model defense and application privacy.
The defenses operate on two tracks. For individual API requests, OpenAI combines targeted model training with input and output filters for messages and tool calls, plus API-level blocks that prevent older models from accessing protected reasoning. Against coordinated campaigns, the company bans thousands to tens of thousands of accounts weekly. The updated safeguards specifically target adaptive black-box red teaming methods that attempt to reconstruct how the model processes information during generation. By filtering tool calls and messages at the API boundary, OpenAI prevents extraction prompts from reaching the reasoning pipeline. The account bans serve as a deterrent against large-scale distillation campaigns that rely on distributing attack traffic across many identities.
Anthropic is pursuing a similar path, reportedly succeeding in 90 to 99 percent of cases by banning tens of thousands of accounts to curb systematic distillation attacks. Both companies are pushing for a coordinated industry and policy response to these extraction methods, highlighting the growing operational cost of defending frontier models against automated reasoning theft. The move signals a shift in how labs manage model security, prioritizing account-level enforcement and API gating alongside model-level training adjustments.
Signal Blocks Screenshots to Counter OS-Level AI Agents
Signal is restricting screenshot capabilities on Android and iOS to prevent operating system-level AI agents from reading private messages. The change, reported by TechCrunch and The Verge, blocks AI assistants like Siri and Gemini from capturing on-screen content. The restriction applies specifically to iOS 27 and Android builds. Users can disable the restriction in settings, but Signal warns this breaks the app’s privacy guarantees.
Meredith Whittaker, Signal’s president, stated that OS-level AI agents are scanning user data outside the app’s control. > "We want to be very clear: it is a serious problem that Signal – and application developers beyond – lack robust tools to prevent AI agents from undermining application-level privacy," Whittaker wrote.
Illustrative image: theaflowers / Pixabay
The screenshot block addresses one vector but leaves others open. AI agents can still access application data through device file systems, and Signal admits the fix is incomplete. The company is also unable to implement equivalent protections on macOS due to missing developer tools from Apple. Signal is requesting that Apple, Google, and Microsoft provide simple technical controls for app developers to restrict AI data access without degrading core user experience.
This follows Signal's previous move to block screenshots on Windows to counter Microsoft's Recall feature. The app notes that third-party agents, such as Meta's Muse, could use deep accessibility settings to bypass application-level privacy controls. Signal emphasizes that the current integration of AI agents into operating systems, without common-sense remedies, represents a structural threat to application-layer privacy. The company is waiting for OS providers to publish accurate documentation and technical tools before it can build a more comprehensive defense.
The shift highlights a growing friction between operating system providers and independent developers. OS vendors are embedding agentic AI capabilities that scan screen content to learn about user behavior, a design choice that directly conflicts with end-to-end encryption guarantees. Signal argues that without standardized APIs to block AI data access, developers are forced into binary choices: degrade the user experience by blocking core OS features, or accept that private conversations are exposed to local databases and remote servers controlled by third parties. Signal is positioning itself as an early indicator of this ecosystem-wide conflict, urging the industry to treat application-layer privacy as a foundational requirement rather than an afterthought.
Signal’s request for OS-level tooling underscores a broader developer experience gap. Current accessibility frameworks allow AI agents to read screen content with minimal friction, but there is no equivalent standard for denying that access on a per-application basis. Without a standardized deny-list or permission layer, developers must rely on workarounds like screenshot blocking, which degrade usability and fail to stop background data collection. Signal is calling for Apple, Google, and Microsoft to expose simple technical interfaces that let apps declare which data streams should be off-limits to system AI.
Sources
- techcrunch.com: Signal restricts mobile screenshots as it confronts ‘existential threat’ of AI surveillance | TechCrunch
- the-decoder.com: OpenAI makes it much harder for attackers to extract how its latest AI models think
- theverge.com: Signal is blocking Gemini and Siri AI from reading your screen
PNEUMETRON EDITORIAL
Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier, not to replace real reporting.




