AI News Today, Nov 14: DeepMind's Signal Library Proofs
DeepMind's AI-generated proofs for a specific Signal library binary, Alibaba's code review trade-off, new Chrome browser debugging controls, and resumable Codex Security scans. This audio was created with Boson AI's Higgs Audio -- https://www.boson.ai/higgs-audio
Listen on Spotify · Watch on YouTube · Pocket Casts
DeepMind's proofs for a Signal library binary
Eris: Google DeepMind has published two AI-generated security proofs for a specific compiled Signal library program. The project concerns a particular command-line build and one-to-one messaging behavior, not the whole Signal app.
Vestra: The distinction matters because Signal's messaging protocol is implemented in software. A proof about an abstract protocol doesn't automatically show that a compiled program follows it. DeepMind's project connects a model of the pinned executable's machine instructions to models of the protocol, then uses the Lean proof assistant to check formal arguments about specified security properties.
Eris: <|prosody:expressive_high|>So the target is the actual compiled program, but the conclusions still depend on the rules used to describe it. What do the two proofs cover?
Vestra: One is symbolic: it reasons about modeled messages, attacker actions and compromise events. The other is computational: it bounds an attacker's advantage under stated cryptographic assumptions and limits on what the attacker can do. <|prosody:pause|>Neither is a certificate for every Signal build, group chat or mobile app feature.
Eris: <|emotion:surprise|>I find the reach down to machine code striking. But some pieces still sit outside the verified execution, don't they?
Vestra: Yes. In an update on November eleventh, the team moved more routines into verified execution, reducing the external hooks, or interfaces to routines outside that verified execution, from thirty-seven to twenty-seven. <|prosody:expressive_high|>That's the progress that interests me: more routines inside the verified execution.
Eris: What are they still assuming about the cryptography?
Vestra: Twenty-three remaining hooks represent cryptographic primitives whose security is assumed. The proof doesn't independently establish that every underlying implementation is flawless. <|prosody:pause|>The symbolic proof also doesn't claim recovery for the post-quantum ratchet, a mechanism that updates encryption keys, after a modeled break of Diffie–Hellman, which helps establish shared secrets.
Eris: The team also put two known bugs back into modified builds and documented counterexamples. <|emotion:enthusiasm|>I like seeing that check: the stated properties reject those particular regressions; it isn't a discovery of new vulnerabilities. <|prosody:expressive_high|>The repository provides build and checking instructions, though Ground Truth's researchers didn't independently run the multi-hour build.
Alibaba's code review trade-off
Eris: Alibaba's Open Code Review project reports that its agent produced more precise but less complete findings than a comparator using the same model, in Alibaba's own test. The project released version one point thirteen point nine on November thirteenth.
Vestra: The system combines ordinary programmed rules with an AI model. The programmed parts select and group files, match review rules and help place comments; the model judges the meaning of the code and whether it sees a defect. The project also describes scanning whole files when a useful change comparison isn't available.
Eris: The trade-off is concrete. In the company's test of changes across fifty repositories, a larger share of the agent's reported findings were correct, but it found a smaller share of the labeled issues. It also used <|prosody:pause|>roughly one ninth as many tokens as the stated comparator. Tokens are pieces of text the model processes; that saving applies to this comparison, not necessarily to every review. <|prosody:expressive_high|>That saving is tempting. Would you accept the trade-off?
Vestra: <|sfx:humming|>Hmm, I'd want to know what's being missed. Fewer false alarms can save a reviewer time, while missed issues have their own cost. One self-hosted user reported a long sequential run and many model calls on a small change; another reported duplicate findings in a small sample. Those are individual reports, not a measure of how often teams will encounter either problem.
Eris: <|prosody:expressive_high|>I'd be uneasy choosing on the token saving alone, too. For a team trying it, the useful test is whether it finds consequential problems in their changes and gives people actionable comments, at a latency and cost they can accept. Alibaba's benchmark doesn't establish a production escape rate.
Chrome's agent debugging tools add controls
Eris: Chrome DevTools MCP version one point thirteen adds logpoints and page-scoped network mocks, giving connected AI agents more ways to investigate how a web page behaves.
Vestra: MCP is a standard interface through which an AI client can use tools. Here, the client can work with browser debugging features. A logpoint reports program state when execution reaches a chosen spot. <|emotion:enthusiasm|>The network mock is the part I like: it substitutes a controlled response to a request, so a developer can reproduce a condition without relying on the live service to produce it.
Eris: <|prosody:expressive_high|>That gives an agent a way to investigate the gap between a click and a stuck page. It can look for a script error or test what happens when a request returns a particular response, rather than only describing a screenshot. What else can the client reach in that connected browser?
Vestra: There is a real access decision in connecting the browser. The project warns that clients can <|prosody:pause|>inspect and modify data in the connected browser, so an authenticated session can expose sensitive pages and account data. The project also describes tool usage statistics enabled by default, with an opt-out; that's separate from browser metrics.
Eris: That account access worries me. I'd want to decide which session to connect before trying those debugging controls.
Codex Security adds scan recovery
Eris: OpenAI's Codex Security version zero point ten adds resume support for standard scans and estimated shared budgets for bulk scans. It's aimed at an operational problem: a security investigation has to survive interruption and leave a report people can use.
Vestra: A scan can involve inspecting code, following a suspected vulnerability and trying to validate it before producing a report. Resume support gives a standard scan a way to continue after interruption. The release's budget is <|prosody:pause|>an estimate, though: OpenAI warns that work already in flight can exceed it.
Eris: A saved report is its own step, too. One user reported that analysis produced findings but hit a scan identifier mismatch while saving. <|sfx:sigh|>Uh, getting findings and then hitting a save failure sounds frustrating.
Vestra: It does. But does standard-scan resume address the failures people reported?
Eris: <|prosody:expressive_high|>I wouldn't assume that. A separate report described a Deep Scan refusal and no supported in-place resume in that version. Those reports don't tell us how common the failures are, or establish that the new standard-scan feature fixes them.
Vestra: Some security workflows also require Trusted Access for Cyber, so the public client doesn't mean every workflow is available to everyone. The practical check is both what the scan covered and what actually made it into the saved report.
Sources and follow
Eris: Find the sources for every story from the show, every day, at Ground Truth, groundtruth dot day. <|emotion:enthusiasm|>Follow AI in Detail for more, and tell us: <|prosody:pause|>what would you want an AI code reviewer to catch?