Opinion: Prompt Injection Is a Design Problem, Not a Bug
AI agents follow instructions hidden in what they read. OWASP says it is unclear if fool-proof prevention exists. My view: limit what a hijacked agent can do.

Security for LLMs and AI systems: attacks, defenses and incidents. Recent posts compare prompt injection and jailbreak defense frameworks like LLM Guard and Prompt Guard, cover the OpenCode API key breach, and review LLM red teaming tools such as Garak and PyRIT.
AI agents follow instructions hidden in what they read. OWASP says it is unclear if fool-proof prevention exists. My view: limit what a hijacked agent can do.

Apache credits Anthropic's OSS Scanner with first reporting CVE-2026-119618, a critical HTTP/2 flaw fixed in Apache HTTP Server 2.4.71 on November 11.

VulnCheck counts 442 CVEs credited to Anthropic in Glasswing's first seven months. Two are exploited, and the ledger shows 217 of 37,966 findings fixed.

With stolen Claude credentials, Darktrace steered Claude Code through memory stores and moved a 1 GB file out via the Files API, in a lab test.

The Intercept reports Vovan and Lexus posed as Ukraine's foreign intelligence head on a 17-minute call with three OpenAI staff. OpenAI confirmed it was a prank.

Railway says stolen AI keys ran up an estimated $740,000, and that about 100,000 workspaces is an upper bound of those that could have had variables decrypted.

DeepMind says AlphaProof Code proved key security properties of Signal's libsignal encryption code in about four weeks, not the whole Signal app.

Cisco is adding Claude Mythos to AI Defense red teaming, and IBM plans Mythos code scanning in Concert Protect. Customers get results, not open prompts.

Australia accepted all 18 recommendations of its OpenAI incident review and plans a frontier AI regulator. A criminal offence is only being considered.

Enclave says DeepSeek V4.1 Flash verified 11 of 11 on its AI Hacking Race, but five wins used shortcuts; after closing them, a rerun verified 4 of 11.

A plain guide to how Claude Code, Codex and Copilot CLI decide when an AI coding agent must ask before it edits files, runs commands or goes online.

Fireworks AI says its October breach reached billing data and, for some accounts, secrets, datasets and incomplete models, with evidence of LLMJacking links.

Prosecutors say three men allegedly used AI to mimic CEOs and big-firm lawyers and got staff to wire over $129 million. One defendant denies any role.

Microsoft says attackers use the ARTEX agent framework to find exposed GitHub secrets and reach Azure and Microsoft 365. One case took five minutes.

On Vals AI's CyberOffense ranges, open-weight DeepSeek V4.1 Flash scored 72.7% overall vs 80.9% for OpenAI's restricted GPT-6 Sol Daybreak.

Human-guided AI agents at DeepMind found cheaper attacks on legacy 1024-bit RSA. What ran, what is only estimated, and what it means for 2048-bit keys.

On Nov 6 OpenAI partnered with Hacktron to test its own systems, expanding the firm's Daybreak Red access. What the announcement says and omits.

OpenAI launched GPT-6 Cyber on Nov 5 for approved Daybreak Red customers. What it is, who gets access, listed prices, and what OpenAI claims.

Apiiro found secrets in 1,122 public copies of Claude Code's settings.local.json, saved by the Don't Ask Again button. What happened and what to do.

Anthropic's FlashProve proves which model answered a request, at 70 to 95 times the GPU time. Hacker News asked who needs it and what it costs; X focused on cost and correctness.
