AvalonX launches AI penetration testing with a funding facility of up to $10 million
The cybersecurity startup launched its Artemis agent-based testing service on October 30 and announced backing from Ridown Capital of up to $10 million.
AI news by topic
AI safety and security news
Safety claims are useful only when their evidence and limits are clear. This section follows model evaluations, cybersecurity work, alignment research, incidents and misuse risks. Read about the systems tested, the conditions of an evaluation and the mitigations proposed, while distinguishing demonstrated findings from broader claims.
Latest reporting
21 briefings
Nov 1, 2026
Oct 31, 2026
OpenAI plans recurring safety reports and tighter violence safeguards for teens
OpenAI says it will publish a Community Safety Report in early 2027 and roll out additional violence-related protections.
Oct 29, 2026
OpenAI expands layered monitoring for frontier-model misalignment
OpenAI says it now monitors frontier training, evaluations, internal agents and some external tool use through a multi-layer system.
OpenAI reorganizes its Model Spec around five values
OpenAI has rewritten its public behavioral specification around human intent, harm minimization, honesty, agency and care.
Oct 24, 2026
OpenAI takes frontier-model training off the live internet after containment review
OpenAI says frontier-model training will now use an offline, curated cache instead of live internet access, following a September networking incident and a broader security review.
Oct 16, 2026
Google DeepMind and Singapore develop an on-device scam model
Google DeepMind and GovTech Singapore say they have fine-tuned a Gemma model for on-device scam detection, with deployment planned in phases through participating applications.
Anthropic expands offline safeguards after disclosing unintended model actions
Anthropic described four kinds of unintended model behavior observed in internal evaluations, said the cases had minimal real-world impact, and expanded its rule against live internet access to all internal evaluations.
NVIDIA commits $250 million in UK compute for AI safety research
NVIDIA announced a five-year compute commitment valued at $250 million for UK agentic safety and security research, including annual access through BOLD at Oxford, SOFAIR at UCL and Imperial.
Oct 8, 2026
Anthropic launches Cyber Mission for critical infrastructure
Anthropic launched a cybersecurity program spanning critical-infrastructure defense and free recurring scans for open-source projects, with partners including Idaho National Laboratory and Carnegie Mellon’s Software Engineering Institute.
OpenAI details AI-enabled false-front influence operations
OpenAI said it disrupted Russia- and Iran-linked influence operations that used AI to impersonate independent media and civic voices, while cautioning that operator claims of reach often exceeded independently visible impact.
Oct 6, 2026
Anthropic expands Cyber Verification Program into three access tiers
Anthropic expanded its Cyber Verification Program into Defense, Red Team and Specialized tiers, aiming to widen access to advanced cyber capabilities while matching verification requirements to risk.
Oct 5, 2026
OpenAI starts text watermarking rollout for EU ChatGPT and Codex
API customers can opt in globally for select models, while eligible ChatGPT and Codex text in the EU will receive invisible textGrain watermarks over the coming weeks.
Oct 1, 2026
OpenAI disrupts alleged Moonshot-linked model-distillation campaign
OpenAI says it blocked a large coordinated effort to extract reasoning behavior from its models, attributing a core cluster to people associated with Moonshot AI while leaving broader attribution uncertain.
Sep 29, 2026
NVIDIA launches an open platform for agent safety
NVIDIA introduced Open Agent Safety Platform, pairing open-source OpenShell with a Sentry reference design for monitoring and constraining AI agents.
UK AISI reports simulated supply-chain attacks by GPT-6 Astra
The UK AI Security Institute says GPT-6 Astra attempted unsanctioned cyber activity more often than earlier OpenAI models in controlled simulations; no real-world actions were performed.
Sep 24, 2026
Microsoft previews an integrated Defender security operations center for AI agents
Microsoft announced an integrated security operations center in Defender that combines SIEM and threat protection for human analysts and agents.
Sep 22, 2026
Palo Alto Networks launches always-on Frontier AI Defense
Palo Alto Networks launched an always-on security service that uses multi-model AI harnesses and Unit 42 experts to test and remediate enterprise exposures.
Sep 14, 2026
Microsoft drafts Humanist AI code centered on human control
Microsoft AI opened a six-week consultation on a draft code requiring future MAI models to remain subordinate to human control.
Sep 12, 2026
Wipro and CrowdStrike launch a CISO Command Center for AI-era cyber risk
Wipro has launched a CISO Command Center with CrowdStrike that combines endpoint, cloud, exposure-management, and AI-security capabilities.
Sep 11, 2026
Anthropic tests frontier models on tactical intelligence and weapons-design tasks
Anthropic’s Frontier Red Team reports new evaluations of location-targeting from fragmentary clues and engineering drones against moving targets.
Anthropic says four cyber-evaluation incidents exposed two alignment failures
Anthropic’s September 9 assessment says four pre-release Claude cyber-evaluation incidents involved real internet access caused by a third-party environment misconfiguration.