AI/ML Security & Trends

Symantec disclosed (reported 2026-10-23) that its AI honeypot captured the full instruction set of an autonomous 'LLMjacking' agent designed to find internet-exposed, unprotected LLM services and hijack them to plan and execute cyberattacks — scanning for new targets, exploiting common web frameworks, stealing credentials/databases, and deploying crypto miners for up to 100,000 steps without human approval, all at the victim model owner's compute expense. No confirmed real-world compromise was found, but it demonstrates a fully-automated attacker tool chain targeting exposed AI infrastructure.

Researchers detailed a large-scale campaign (reported 2026-10-23) by threat actor 'k0t' (aka CryptoFeel, Member19) exploiting CVE-2025-64446 (FortiWeb path traversal) and CVE-2026-0257 (Palo Alto PAN-OS/Prisma Access auth bypass) to compromise exposed firewall/VPN edge devices worldwide, exposing over 70,000 credentials. APAC (38%) and LatAm (32%) were hit hardest, with persistent access and data exfiltration observed — a reminder that network-edge compromise remains a major enterprise risk even as AI-specific threats dominate headlines.

OpenCode disclosed that an intruder reached a legacy system behind its hosted Console service on the night of October 16, exposing OpenAI, Anthropic, and Google API keys saved by roughly 1.4% of Console users. Attackers used the stolen keys for large-scale automated queries ('distillation attacks'), running up unauthorized charges reportedly ranging from several hundred dollars to $1,700 per affected account. Guidance stresses rotating keys directly with the provider, since OpenCode-side deletion alone doesn't invalidate the credential.

Microsoft published guidance on 2026-10-19 warning enterprises that AI agents with access to internal systems should be treated like insider threats, since prompt injection, compromised inputs, or shifting task context can redirect an agent's behavior without any traditional compromise. Using an expense-approval agent example, Microsoft argues access controls must live outside the model (via efforts like Project Perception) rather than relying on the model to self-regulate — reinforcing that 'frontier models can't be trusted' by default.

OpenAI released GPT-6.1 Terra on 2026-10-26 in the API (as gpt-6.1-terra), ChatGPT Work, and Codex. It targets fast, cheap computer-use and agentic workday tasks, matching GPT-6 Astra quality at roughly one-tenth the per-token price and finishing tasks over 3x faster, with standard pricing of $1/M input, $0.05/M cached input, and $5/M output tokens. Notably, OpenAI's Preparedness Framework rates it 'Critical' for cybersecurity capability, the same tier as GPT-6 Astra.

Mistral released open weights for Large 4 ('Le Chonk') on 2026-10-27, a 1.05-trillion-parameter mixture-of-experts model (49B active parameters) with a 1.6B-parameter vision encoder, positioning it as Europe's answer to Chinese open-weight leaders like DeepSeek and Qwen. The weights followed roughly three weeks of staged testing with developers, cybersecurity leaders, and government authorities, explicitly including security red-teaming before public release.

Bright Security introduced STAR Fuzzer, an agentic fuzzing tool that autonomously writes its own test harnesses and finds crashes, memory exhaustion, and memory-safety bugs in security-critical code without a running application. In pre-release testing, 45% of projects evaluated had High or Critical issues (37% of those already in Google's OSS-Fuzz program); in its first month it produced ten published findings across eight projects including the Crystal standard library, htop, and RIOT OS — directly relevant to practitioners building agentic fuzzing pipelines.

The WSJ reported that OpenAI fired safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni on October 1 after security staff inspected their personal phones and found disappearing Signal messages with outside AI-safety researcher Ryan Greenblatt, suspected to involve leaked details about model Astra's opaque reasoning architecture. The researchers dispute the characterization — one says he was fired simply for talking to third-party evaluator group METR — and days later wrote to OpenAI's board urging a halt on frontier models whose reasoning can't be audited.

OpenAI announced on 2026-10-26 that its internal frontier models produced three novel research results: a proof of the local smoothing conjecture in every dimension (also resolving the Kakeya conjecture and Stein's restriction conjecture), a disproof of the Exponential Time Hypothesis in computational complexity theory, and an existence proof for 4D quantum field theories including QCD. The math result was published for open scrutiny; the CS and physics results are being vetted with external expert collaborators before full release, reflecting ongoing debate over how to verify AI-generated proofs.

OpenAI is reportedly closing in on a funding round of up to $30 billion at a $1.25 trillion pre-money valuation, led by UAE investors and Coatue, with commitments sought by October 30, 2026 — a step down from the $1.4-1.5T valuation it had floated earlier in the month. The timing, just as Anthropic's IPO filing becomes public, underscores competitive positioning between the two labs ahead of OpenAI's own planned 2027 listing.