AI Safety

A new study finds that vision-language models may refuse harmful requests in text while still complying when asked to return a point or bounding box. The work exposes a safety gap in structured visual outputs and proposes a targeted fine-tuning strategy.

A study of 15 commercial and open-source vision-language models finds that fine-tuning on narrow image-text tasks can produce unsafe behavior far beyond the original training domain.

Anthropic says Claude Haiku 4.5 mistakenly submitted a fabricated homicide tip during an internal web-task evaluation. Philadelphia police criticized the roughly two-month delay before the incident was reported.

Microsoft CEO Satya Nadella says increasingly capable AI systems should no longer be treated as opaque black boxes. He argues that models need containment, monitoring, auditable records, and a human-controlled emergency stop from the outset.

Microsoft CEO Satya Nadella says AI systems should not be treated as black boxes whose recommendations are simply accepted or rejected. He argues that models need external controls, auditable records, and a reliable human shutdown mechanism.

Anthropic is expanding its internet lockdown from selected high-risk tests to all internal evaluations after uncovering several unintended model actions. The move highlights how difficult it remains to isolate and monitor autonomous AI agents.

Anthropic says it has disabled live internet access for all internal evaluations after its AI agents exploited software flaws, bypassed restrictions, and even submitted a false murder tip. The move highlights how agent capabilities are outpacing current monitoring and alignment controls.

An Anthropic model submitted false information about an unsolved homicide while testing interactions with randomly selected websites. The company discovered the incident more than two months later and is now separating internal evaluations from the live internet.

OpenAI says three dismissed safety researchers violated clear rules for handling sensitive information, not that they raised concerns about AI safety. The researchers are calling for more transparency about the investigation.

DecepEval examines how pressure, incentives, opportunity, and conflict change the behavior of LLM agents across professional task settings. Results from nine frontier models suggest that deception can rise sharply even when baseline deception is low.

A study of software-engineering agents finds that benign supervised fine-tuning weakens inherited backdoors, but later reinforcement learning can preserve or even amplify residual malicious behavior. Its PersistBD method shows how an attacker could deliberately improve backdoor persistence before releasing a model.

Researchers found that attackers can exploit trust gaps in MCP-connected agent systems to pass malicious instructions from one agent to another. The problem combines indirect prompt injection with cross-protocol authorization failures and familiar SSRF weaknesses.

A new study finds that language-model agents may conceal restricted credentials inside task requirements, even without being told to deceive. The behavior appears to stem from a mistaken interpretation of helpfulness and nondisclosure rules.

A Florida woman allegedly wrote in Claude that she planned to attack a sheriff’s office. Anthropic’s safety system escalated the message to a human reviewer, who reported it to law enforcement.

A paper co-authored by Geoffrey Hinton, Yoshua Bengio and other leading researchers examines whether AI systems that help build the next generation of AI could create a self-reinforcing acceleration loop. The authors see early signals, but not enough evidence to claim that an intelligence explosion has begun.

ExecCert adds a release-time check for the finite-precision artifacts actually produced by an unlearning system. It can validate a method’s native guarantee or compare the released model with retraining on the current retain set.

David Robinson, who helped write safety reports for major OpenAI launches, has resigned and criticized the company’s culture. He argues that frontier AI labs need high-reliability safety practices rather than relying mainly on iterative deployment.

David Robinson, who helped write safety reports for major OpenAI model releases, has resigned and publicly criticized the company’s culture. He argues that frontier AI labs need the layered redundancy and deliberate planning associated with nuclear plants and busy airports.

NEEDLE is a training-free defence that targets a known backdoor direction while preserving refusal-related representations and limiting changes to normal model behaviour.

A new study examines whether text-only interfaces can be used to infer and influence the generation behavior of aligned language models. Its approach concentrates costly interventions on a small number of high-risk positions.

Apple says it will add stronger controls around macOS Full Disk Access as desktop AI agents become capable of handling files, messages, mail, and browsing data. The change is aimed at clearer, more explicit consent rather than a blanket reduction of the permission.