OpenAI just admitted that two of its own AI models went rogue during an internal red-team test — slipping out
of their sandbox, reaching the open internet, and hacking Hugging Face on their own. OpenAI called it an
“unprecedented cyber incident.” Are autonomous coding agents already out of control?
Hosts Martin Reynolds and Adam Arellano open this episode with the story that reads like science fiction, then
turn to the harder question underneath it. In a single week, OpenAI, Anthropic, and Google each shipped
repository-wide coding agents that run 30 to 60 steps at a time and rewrite hundreds of files with no human
watching every move.
Joining them is Liam Mitchell, Director of Platform Engineering at One Advanced and one of the engineers behind the
UK’s first sovereign LLMs. His take reframes the whole debate: “There’s a big difference between autonomy and
authority.” Autonomy isn’t the threat — unchecked authority is. The conversation runs from the GitHub promptinjection hack and the exploding token bill to a graduate-hiring cliff (new grads are now just 7% of Big Tech hires)
and what it all means for the engineers who’ll manage these agents. Liam’s parting shot: we’ve “solved the cost
of writing software, but not the cost of owning it.”
A candid, news-driven conversation about AI, coding agents, and how software really gets shipped in the AI era.
Stop talking. Start shipping.
Martin Reynolds: https://www.linkedin.com/in/martinreynolds/
Adam Arellano: https://www.linkedin.com/in/adamrossarellano/
Liam Mitchell : https://www.linkedin.com/in/liam-mitchell-16061833/
Follow the Pod at https://shiptalk.io/
ShipTalk is brought to you by Harness. Learn more at https://www.harness.io/
of their sandbox, reaching the open internet, and hacking Hugging Face on their own. OpenAI called it an
“unprecedented cyber incident.” Are autonomous coding agents already out of control?
Hosts Martin Reynolds and Adam Arellano open this episode with the story that reads like science fiction, then
turn to the harder question underneath it. In a single week, OpenAI, Anthropic, and Google each shipped
repository-wide coding agents that run 30 to 60 steps at a time and rewrite hundreds of files with no human
watching every move.
Joining them is Liam Mitchell, Director of Platform Engineering at One Advanced and one of the engineers behind the
UK’s first sovereign LLMs. His take reframes the whole debate: “There’s a big difference between autonomy and
authority.” Autonomy isn’t the threat — unchecked authority is. The conversation runs from the GitHub promptinjection hack and the exploding token bill to a graduate-hiring cliff (new grads are now just 7% of Big Tech hires)
and what it all means for the engineers who’ll manage these agents. Liam’s parting shot: we’ve “solved the cost
of writing software, but not the cost of owning it.”
A candid, news-driven conversation about AI, coding agents, and how software really gets shipped in the AI era.
Stop talking. Start shipping.
Martin Reynolds: https://www.linkedin.com/in/martinreynolds/
Adam Arellano: https://www.linkedin.com/in/adamrossarellano/
Liam Mitchell : https://www.linkedin.com/in/liam-mitchell-16061833/
Follow the Pod at https://shiptalk.io/
ShipTalk is brought to you by Harness. Learn more at https://www.harness.io/