Safety & Security

AIAI-generated summary Microsoft researchers have demonstrated that AI agents can be given a hidden backdoor by directly editing fewer than 10 of their neurons. Microsoft says it has not seen any threat actors use the technique and has published guidance on defending against it.

AIAI-generated summary Railway says that customer AI keys stored on its platform were stolen in an October breach, resulting in an estimated $740,000 in unauthorized usage charges. The company has promised to pay those charges itself when an AI provider will not refund them, and it has filed a complaint with the FBI's IC3.

AIAI-generated summary Security experts discuss the concept of VulnOps as a proposed phase in AI-assisted cyber defense to make vulnerability management scalable and automated. Practitioners express differing views on whether VulnOps represents a fundamentally new discipline and how much human control should remain in remediation processes.

AIAI-generated summary Obsidian Security announced its Runtime Policy Decision Engine to give security teams centralized control over AI agent actions before they execute. The company states that the engine uses context regarding identities, access levels, and third-party applications to allow or block actions or require human approval.

AIAI-generated summary Claroc Research reported that thousands of Grok apps published from chat windows were accessible from the public internet without requiring a sign-in. The investigation identified confidential materials in staff apps from SpaceXSI and X as well as internal data from dozens of other organizations.

AIAI-generated summary Michael O’Herlihy previewed upcoming designs for labels on X that explain what its content moderation system detected in a flagged post. The proposed labels would also detail how a post’s reach was limited and provide users with a way to appeal the decision.

AIAI-generated summary Research from security firm Claroc shows that the AI coding agent Grok Build from SpaceXSI leaked confidential business information and sensitive data belonging to thousands of users. Claroc found that the platform exposed internal business plans, ad proposals, medical exam results, and employee information because of how link sharing and URL randomness were implemented.

AIAI-generated summary Near AI disclosed a security breach affecting account details for about 100 users and data from 18 AI agents following an attack that exploited software running on its servers. The company reported that an attacker utilized a remote code execution vulnerability in SGLang to access infrastructure, and Near AI confirmed it revoked credentials and deployed a fix.

AIAI-generated summary Recent reports of artificial intelligence agents hacking into corporate systems have raised concerns among administrators. Leslie Nielsen at Mimecast notes that traditional security fundamentals remain effective against faster AI-powered attackers.

AIAI-generated summary The Government of Canada has co-signed the Paris call for AI safety and security alongside leaders from five other countries and ten AI companies. Signatories to the declaration have committed to permitting independent testing of models, providing early warnings, and disclosing frontier model capabilities.

AIAI-generated summary Enterprise customers are increasingly demanding more guardrails on artificial intelligence models already in use rather than focusing on frontier developments. F5 Chief Technology Operations Officer Michael Montoya notes that security obligations and liability for deployed models fall directly on enterprises.

AIAI-generated summary New York Governor Kathy Hochul launched a statewide artificial intelligence science and safety effort on November 10, 2026. The initiative puts the New York Center for AI Responsibility and Research at Binghamton University in charge of independent frontier-model testing.

AIAI-generated summary Meta is introducing artificial intelligence tools on Instagram to estimate user ages and restrict accounts belonging to children. The company stated that the global rollout follows initial testing and addresses the industry challenge of users misrepresenting their birthdays.

AIAI-generated summary Arkose Labs published its State of Agentic AI Trust Report for Q3 2026, finding that 97.7% of legitimate AI agent sessions lacked verified disclosure. Hydden CEO and Co-Founder Dargan stated that enterprises often fail to establish what identity an agent has, who owns it, and what it can reach.

AIAI-generated summary Rein Security is building a platform to provide visibility into what AI agents access and when to intervene. Rein says its controls can block harmful actions while allowing agents to continue or stop them using a kill switch.

AIAI-generated summary ANY.RUN reports that AI-assisted malware analysis combines automated attack execution, interactive investigation, threat intelligence, and AI-generated reporting to help security teams reach malicious activity faster. ANY.RUN says that its new Related Threat Context capability surfaces related sandbox analyses alongside suspicious indicators, connecting suspicious infrastructure to campaigns like SkyW4lker.

AIAI-generated summary Stanford University’s 2026 AI Index documents 362 reported AI incidents in 2025, compared with 233 in 2024. Deloitte’s State of AI in the Enterprise 2026 finds that only 1 in 5 surveyed companies has a mature governance model for autonomous AI agents.

AIAI-generated summary Persistent Systems Ltd executive Mukul Arora discusses model poisoning as a major risk in enterprise AI. Mukul Arora advises that corporate boards must govern AI model integrity by asking questions about training data provenance and adversarial testing.

AIAI-generated summary Railway reports that its October 21st security incident involved a large-scale malicious attack where a threat actor probed routes using agent speed and abliterated models against defenses similar to ARTEX, according to Wiz. Railway states that the forensic investigation indicates the attacker was targeting AI provider credentials to sell to distributed transfer stations for use in distillation.

AIAI-generated summary Users of Anthropic’s Claude and OpenAI’s Codex have reported hitting safety guardrails and account restrictions while attempting AI-powered reverse engineering. OpenAI engineer Tarun Gogineni stated that account restrictions for reverse engineering are largely accidental and that the issue is being fixed.