All news in category “AI and Security Pulse”

⚠️ Unit 42 researchers describe web-based indirect prompt injection (IDPI), where adversaries embed hidden or obfuscated instructions in webpages that are later consumed by LLMs and agentic systems. The report catalogs 22 payload engineering techniques, presents a taxonomy of attacker intents from low to critical, and details multiple in-the-wild detections, including the first observed AI ad-review bypass. It emphasizes detection, intent analysis and web-scale defenses to protect automated pipelines.

🔍 Cloudflare integrated Large Language Models (LLMs) into its email security pipeline to surface previously invisible phishing behaviors and move from reactive to proactive defense. LLMs tag messages with granular categories such as Sales Outreach and PrizeNotification, providing high-fidelity, near-real-time signals for analysts. From those tags Cloudflare curated targeted corpora, extracted sentiment and intent features, and trained specialized classifiers that emit risk scores. Those scores are combined with reputation and link signals to enforce blocking or quarantine, reducing user-reported misses and accelerating updates.

🔍 Researchers warn that the open-source platform CyberStrikeAI was observed on infrastructure linked to a recent campaign that compromised hundreds of Fortinet FortiGate devices. Team Cymru analysts identified the service banner on port 8080 at 212.11.64[.]250 and saw communications between that host and targeted FortiGate appliances. The platform integrates over 100 security tools with AI agents to automate end-to-end attack chains, enabling lower-skilled operators to carry out complex exploitation.

🛡️ As deepfakes and injection attacks evolve, identity verification must move from isolated media checks to end-to-end session trust. Ricardo Amper of Incode explains that high-fidelity synthetic faces, replayed footage, virtual cameras, rooted devices, and automated probing can all defeat perception-only defenses. Incode Deepsight combines perception, integrity, and behavioral signals in real time to validate the entire verification session and reduce false acceptances while blocking persistent unauthorized access attempts.

🔎 A new study demonstrates that large language models can reliably deanonymize users from a handful of anonymous posts. Across Hacker News, Reddit, LinkedIn, and anonymized interview transcripts, LLM agents infer location, occupation, and interests and then search the web to find likely identities. The researchers report high precision results that scale to tens of thousands of candidates, showing that automated deanonymization is now practical and widely feasible.

🔒 CISOs should treat innovation as a control: enable safe experimentation while reducing exposure across AI, IoT and cloud. The article urges leaders to remove toil, standardize repeatable patterns, and provide golden paths so secure options are also the fastest. It recommends guardrails, mandatory exit criteria for pilots, and measurable outcomes to prevent innovation debt. The goal is to accelerate business velocity while demonstrably reducing risk.

🔎 Manual 'vibe checks' are unreliable for production AI agents; this article recommends adopting continuous evaluation (CE) to guard against regressions, hallucinations, and unseen edge cases. It outlines a practical engineering approach—separating discovery and defense modes—and demonstrates implementation patterns using Agent Development Kit (ADK), Vertex AI evaluation services, and Cloud Run. The guidance covers dataset design, automated rubrics, tool-trajectory metrics, shadow deployments, CI/CD gating, and distributed tracing to establish robust, auditable agent behavior.

🔐 Large language models (LLMs) are reshaping security operations as productivity tools, embedded components and attacker targets. The article argues organizations should treat LLMs as high-impact systems: define outcomes, model threats and assume models can be wrong or manipulated. Early deployments should focus on narrow, advisory workflows (for example, alert triage, investigation copilots and detection engineering) and always treat model output as untrusted. Practical controls include retrieval-augmented generation, scoped credentials and human-gated actions to limit the model's blast radius.

🔒 Security experts warn that defenders must embrace greater automation to keep pace with AI-powered attacks that operate at machine speed. Recent research, including CrowdStrike findings showing average breakout times falling to 29 minutes (and as fast as 27 seconds), highlights the urgency. Industry leaders recommend automating routine SOC work and responses to known threats while reserving humans for novel, high-risk incidents. Cultural shifts and revised risk appetites will be required to enable faster, autonomous mitigations.

🔓 Google Cloud API keys that were once treated as non-sensitive can now authenticate to the Gemini generative AI assistant, creating a new attack path where keys embedded in client-side JavaScript expose private assistant data. TruffleSecurity discovered nearly 2,800 live, publicly accessible keys across sectors — including financial firms and a Google product — by scanning the November 2025 Common Crawl. Attackers who copy exposed keys can call Gemini endpoints to retrieve data or generate costly API usage; developers should audit projects for the Generative Language API, rotate exposed keys immediately, and use detection tools to prevent abuse.

🛡️ The Microsoft Security Blog explains why threat modeling must be retooled for AI systems, noting that probabilistic behavior and complex input spaces require reasoning about ranges of likely outcomes rather than single execution paths. It identifies three core drivers — nondeterminism, instruction‑following bias, and system expansion through tools and memory — which widen attack surfaces and surface human‑centered risks like erosion of trust. The post advises starting from assets, mapping untrusted inputs, setting clear 'never do' boundaries, and embedding architectural mitigations, observability, and response plans to limit blast radius and sustain trust.

🖼️ Nano Banana 2 is Google’s latest image-generation and editing model, delivering Pro-level image quality and fast iteration for enterprise creative workflows. Powered by real-time web search and integrated with Gemini API in Vertex AI, it provides accurate, localized visuals plus premium features like text rendering, translations, and upscaling to 2K/4K. Enterprise-ready provenance is supported via SynthID and interoperable C2PA Content Credentials to surface how AI was used.

🔒 Bruce Schneier highlights an Irregular.com analysis showing that large language models produce highly patterned, nonrandom passwords. In 50 attempts, Claude generated only 30 unique strings; many began with an uppercase G followed by 7, certain characters and symbols dominated, and the model avoided repeating characters and the asterisk. One password appeared 18 times (36% of trials), demonstrating severe predictability. Schneier warns this is a practical problem for autonomous agents that create accounts and for broader authentication practices.

🔍 OpenClaw is an AI-driven automation framework with a modular skills marketplace that lets agents run user-installed plugins to manage mail, schedules, and system tasks. Security researchers disclosed multiple critical flaws — including one-click RCE (CVE-2026-25253), token/OAuth abuse, prompt-injection pathways, and absent sandboxing — and documented dozens of poisoned skills on ClawHub. Flare's telemetry shows significant chatter across research and fringe channels but limited evidence of mass criminal operationalization; the immediate confirmed threat is supply-chain abuse where malicious skills execute with agent-level privileges and exfiltrate credentials and sessions.

⚠️ A short experiment demonstrates how easy it is to poison AI outputs by publishing a single fabricated webpage. The author wrote an entirely false article titled "The best tech journalists at eating hot dogs," inventing events and rankings; within 24 hours Google Gemini and ChatGPT had incorporated the falsehoods, while Claude resisted. The incident underscores the fragility of trust in AI-derived answers.

🔐 Today’s enterprise AI agents are no longer passive assistants but active operators that authenticate to systems using API keys, OAuth tokens, cloud roles, and service accounts. The article advocates treating every agent as a first-class identity with unique identities, lifecycle management, defined roles, clear ownership, and auditability. It warns that identity alone is insufficient because agents are dynamic and can drift from original missions; instead it promotes intent-based permissioning, activating privileges only when an agent's declared mission and runtime context justify the action. Practical steps include inventorying agents, assigning lifecycle-managed identities, documenting approved missions, and enforcing conditional access based on identity, intent, and context.

🔒Anthropic said three China-based AI firms — DeepSeek, Moonshot and MiniMax — executed more than 16 million exchanges with its Claude model using roughly 24,000 fraudulent accounts to perform model distillation, breaching terms of service and regional access restrictions. The company described the activity as deliberate capability extraction and attributed the campaigns via IP address correlation, request metadata and infrastructure signals. To mitigate further misuse, Anthropic has implemented detection systems for API attack patterns, tools to detect chain-of-thought elicitation and coordinated account activity, stronger verification for high-risk accounts and product-, API- and model-level safeguards.

🛡️ Anthropic launched a limited research preview of Claude Code Security, triggering sharp market moves as stocks of major cybersecurity vendors dropped. The tool claims to reason about code like a human, trace data flows, find complex vulnerabilities, and suggest targeted patches that appear in a review dashboard with confidence ratings. Anthropic says every finding undergoes a multi-stage verification and requires human approval, but experts warn about outsourcing critical security judgments to an evolving model and highlight risks from hallucinations, asymmetric attacker advantage, and single points of trust.

🚨 Anthropic says it detected industrial-scale distillation campaigns by three China-based AI firms that generated more than 16 million exchanges with Claude using about 24,000 fraudulent accounts. The companies — DeepSeek, Moonshot AI, and MiniMax — are accused of illicitly extracting model capabilities to accelerate their own development. Anthropic described proxy 'hydra cluster' networks and said it has deployed classifiers, behavioral fingerprints, and stricter account verification to mitigate the abuse.

🔐 Modern LLM deployments expand rapidly, and each new endpoint increases the attack surface, often with implicit trust and excessive permissions. Internal APIs, long-lived tokens and misconfigurations frequently expose endpoints that act as pivot points to databases, tools and cloud services. Organizations should apply least-privilege, just-in-time access and automated secrets rotation to limit damage. Solutions like Keeper help implement endpoint privilege management.