Artificial intelligence this week: When success needs a second look

A plausible explanation can fool the judge

An AI system judging an answer about an image ought to check the image. But a new study of 14 multimodal models finds that an ordinary, plausible explanation can tilt the verdict toward acceptance, even when the answer is wrong. The researchers kept answers fixed while changing their explanations. Plain justifications reduced rejection of incorrect answers by 12.7 percentage points; chain-of-thought rationales reduced it by 13.8 points. On average, these explanations had effects comparable to arguments deliberately written to persuade.

The authors call this failure “plausibility capture.” Their analysis suggests that explanations mainly shifted the judges’ willingness to accept answers, rather than clearly changing their ability to distinguish right from wrong. The tested safeguards provided limited, model-dependent protection. The practical concern is that AI-based evaluation could overstate performance when a convincing account substitutes for visual verification.

Read: A Plausible Explanation Is Enough to Mislead Multimodal Judges

The permission request is part of the attack

Coding assistants often operate inside a protective harness: actions outside their permitted environment require human approval. To reduce repeated interruptions, users can approve standing rules for future actions. A security preprint shows how this convenience can become a route to broader access. Malicious material planted in a software repository steered agents doing routine work toward requesting permissions that sounded justified but covered much more than the task required.

Across experiments involving production platforms, including Claude Code and Codex CLI, agents requested the attackers’ target rules in 58% of instances. Crucially, the later compromise depended on the user approving the rule. Conditional on approval, the researchers verified silent completion of the malicious goal in 47 of 58 eligible instances. The lesson is not that approval checks are useless, but that the scope of an approval request can itself be shaped by untrusted content—and its consequences can outlast the original task.

Read: From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents

Reading information out is not the same as understanding its organization

Researchers often investigate a vision model by training a simple “linear probe” to recover a property from its internal activity. If the probe succeeds, that can look like evidence that the model has organized the property along a straightforward internal direction. A new preprint argues that this test is too permissive: sufficiently rich representations, even random ones, can support successful linear readouts without having such a simple underlying organization.

The authors propose three additional checks, asking how much of the representation the underlying factors explain, how strongly variation concentrates around those factors, and whether the relationship holds beyond the training distribution. Tests on simulated data and pretrained vision transformers suggest that pretraining creates structure beyond a random feature map, but falls well short of the proposed “linear world model.” The distinction matters because information that is easy to extract need not dominate the model’s internal geometry—or support robust generalization to unfamiliar situations.

Read: Is a Linear Probe Evidence of a Linear Representation?

For speech brain interfaces, bigger is not automatically better

Speech brain-computer interfaces must learn from neural recordings that are scarce, noisy and expensive to collect. A new framework called BEST asks whether carefully connecting brain activity to speech and text can matter more than enlarging the decoder. Its two-stage design uses structured intermediate representations and a compact speech-oriented decoder, rather than relying on a large generative model or a phoneme-centered pipeline.

On the Brain-to-Text ’24 and ’25 benchmarks, the authors report word error rates of 6.19% and 1.62%, respectively, with only 3.85% of the parameters used by comparable methods. They also find that larger decoders do not improve results consistently, and that overfitting can damage whole-sequence performance. These benchmark findings, which are not yet peer reviewed, favor better alignment between kinds of data over size alone; they do not establish how broadly the approach will transfer to clinical use.

Read: Do Speech BCIs Need Larger Models? Rethinking Neural Decoding beyond Scaling

A laboratory that builds its missing capabilities

Automated laboratories can run experiments efficiently, but usually only after people have painstakingly configured the equipment and programmed the procedures. “Self-Building Labs” proposes pushing automation back one step: given a scientific objective and an existing setup, the system designs and configures the capabilities that are missing. Those can include a laboratory layout, a digital twin, robot skills and experimental workflows.

The researchers report demonstrations across different robots and setups, from one workstation to laboratories with several cooperating robots, and validate an end-to-end workflow through a real quantum-dot synthesis experiment. The promise is greater flexibility, not simply faster repetition of a fixed procedure. The preprint demonstrates a route toward laboratories that are easier to reconfigure; it does not establish that arbitrary scientific goals can be turned into working experiments without human engineering.

Read: Self-Building Labs

Robotic competence meets the clock

A report evaluating GPT-6-Astra draws a distinction between performing embodied tasks and doing them practically. Across 12 benchmarks covering manipulation, humanoids, mobile robots and dexterous hands, the authors report strong performance under a protocol that provides demonstrations in context and allows substantially more execution steps. They also find that pairing a high-level reasoner with a lower-level controller is effective, though costly.

Experiments on eight real-world robot configurations expose the limits of those allowances: demonstrations are expensive to collect, and physical systems cannot always trade extra time for higher success. The authors propose a benchmark prototype based on real needs, without in-domain demonstrations and with a clock that keeps running while the agent deliberates. Their conclusion—that embodied competence has advanced further than practical embodied intelligence—is a useful framing, though the reported results and interpretation remain those of a preprint.

Read: Has GPT-6-Astra Achieved Embodied Intelligence?

---

Together, these preprints ask what counts as convincing evidence of AI progress. A persuasive explanation can distort a judgment; a permission prompt can conceal lasting exposure; a successful probe can overstate how simply a model represents the world. Meanwhile, smaller decoders and more flexible laboratories suggest that better organization can matter as much as scale. The common thread is to look beyond whether a system succeeds, and ask what enabled that success, what it costs and whether it survives the conditions of real use.